Now, I would like to address some critical remarks pertaining to CSI. Some people have suggested that the calculation doesn’t really work for the following reasons:
Let’s say that you have a door that opens and closes with random gusts of wind. The open door is marked “state 1" and the closed door is marked “state 0.” After a very long while of marking these states down, you realize that you have a highly complex (in the sense of being highly improbable) and very compressible (specified) pattern of “101010101010101010 ...". This would obviously generate a very high specification, and according to the critic would mean that since we have CSI here, then there must be an intelligent ghost opening and closing the door. Thus, since we know that natural, random gusts of wind can cause a door to open and close, the measurement for specification actually doesn’t work. The critic insists that we have a counter example and thus using a specification as an indication of intelligence is flawed.
Now, before I start, I must remind you that an understanding of the pattern is necessary to calculate its probability, which is essential to calculating a specification. For example, if I see written somewhere the letters “GATTACA,” I could be looking at a short snippet of an actual amino acid sequence or I could merely be looking at the title of a movie. The shannon information content would be either 14 bits or approx. 32.9 bit respectively. In the initial possibility I would be looking at states from a set of 4 potential states and in the latter possibility I would be looking at states from a set of 26 potential states. Thus, based on what the written pattern actually represents, we can have many different probabilities associated to that representational pattern.
So, if we are measuring a representation of a pattern, such as how ACTG can represent adenine, cytosine, thymine, and guanine, on a string of DNA we need to understand the original pattern that we are representing in order to calculate any probabilities.
Now, let’s look at the swinging door example. We have a pattern of 101010101010101010 which represents a swinging door, “1" being open and “0" being closed. Well, what is the probability that we will have a pattern that represents “open door” “closed door” open door” closed door” etc? Well, is there any other possibility? Can we go from “open door” to “open door”? Not unless we have a timer which captures a state every “x” seconds, and in the above scenario that was not included. So no matter how the door is made to swing open and closed we really have a pattern which has a 100% probability, since one state must as a necessity be followed by the other state.
Now, we can measure for specificity and then for a specification. With a probability of 1, we will get a greater than 1 specificity and thus a negative amount of CSI – no specification.
Thus, a proper understanding of the probability of the pattern in question is essential when calculating for a specification.
Monday, February 11, 2008
Specifications (Part I) ... What Exactly are They?
There seems to be much confusion as to what is a specification -- that is, what constitutes CSI? I am no expert on the subject, however, I would like to add my two “sense” based on my understanding of the math and concepts involved. I have started discussing complex specified information here [link]. This is merely an extended discussion, clarification, and a bit deeper of a probe.
First, we must begin by defining a specified pattern. What is the basic idea behind a specified pattern? Why must we define a specified pattern? It is because within patterns, there can be something (potentially non-random characteristics such as function) which separates one set of patterns from all other possible patterns.
Dembski begins to discuss a specified pattern by stating: “The crucial difference between (R) [a random pattern] and (ψR) [a pseudo-random pattern -- the Champernowne sequence] is that (ψR) exhibits a simple, easily described pattern whereas (R) does not. To describe (ψR), it is enough to note that this sequence lists binary numbers in increasing order. By contrast, (R) cannot, so far as we can tell, be described any more simply than by repeating the sequence.” I will continue to discuss the difference between random and pseudo random patterns (as it relates to specificity) further on.
Dr. Dembski describes specified patterns as those patterns which can be described and formulated independent of the event (pattern) in question.
Stephen Myers expounds upon this subject in his paper, “DNA and the Origin of Life: ...”:
... and ...
... and ...
This ability to formulate a pattern independent of the event in question can be understood with a couple of examples. But, we must first start with a communication system. Why? So that we can exchange a pattern across that communication system and see if it can indeed be formulated independent of the pattern in question.
Let’s say that we wanted to send the pattern: 111111111111111111111111111111 across our communication system. First, can we compress it? Yes we can. How? By defining or formulating it differently. On way to do that is by compressing the pattern to “print 1 X 30.” Thus we can send this information across the communication channel and the receiving party will receive the exact same information than if we had sent the original pattern of 1s. Therefore, the original pattern can be defined independently of the pattern itself. So, compressibility is one way to define a specified pattern.
Now, let’s look at another pattern: 1 2 3 5 7 11 13 17 19 23 ... . Is this a specified pattern? In order to evaluate this and the next examples of patterns, we need to understand the difference between what Dembski calls a pseudo-random and a random pattern. A random pattern is algorithmically complex and can not be defined or formulated according to a system of rules whereas a pseudo-random pattern merely looks random, since it is algorithmically complex (not regular as the last example of 1s), however it is not random since it can be defined and formulated as a separate pattern according to a system of rules. This usually entails meaning or function.
So, let’s return to our example of the pattern: 1 2 3 5 7 11 13 17 19 23 ... . Can we define or formulate it in a different way yet retain the same information? Sure we can. This pattern conforms to a system of mathematical rules and can be defined or formulated differently than by merely repeating the pattern. We can send: “all numbers, beginning with one, which are divisible only by themselves and one” or its equivalent mathematical notation across the communication channel and the other end would be able to reconstruct the original pattern -- a sequence of prime numbers.
Now, let’s take a look at proteins. When it comes to measuring specificity, this is exactly like measuring specificity in a meaningful sentence, as I will soon show. Functional specificity merely separates functional pattern “islands” from the sea of random possible patterns. When specific proteins are brought together, you can have a pattern which creates function [http://en.wikipedia.org/wiki/Function_%28biology%29]. [link] That functional pattern itself is formulated by information contained in DNA which is encoded into RNA and decoded into the specific system of functional proteins. The functional pattern as the event in question is defined independently as a pattern of nucleic acids. Thus, the independent formulation for the system of protein function is sent across a communication channel as RNA which is the independent formulation of the function (I have provided a definition of “function” [here “philosophical foudations”] [link]). The RNA is independent of the function itself for which it codes (that is, the information for formulating the function doesn’t come from the protein pattern itself as per Central Dogma of Biology [http://en.wikipedia.org/wiki/Central_dogma_of_molecular_biology] [link]). You don’t send the function itself across the communication channel, you send an informational pattern. Again, that is what is referred to as functional specificity.
What about the pattern: “Can you understand this?” This pattern is most definitely specified, since it can be defined according to a pre-set English dictionary and produce a function through meaning/ideas. How do you pass an idea/meaning through a communication channel? You do so, by using informational units derived from a pre-set system which are separate from the actual pattern of ideas/meaning and which can be processed according to the rules of that pre-set system at the other end of the communication channel. The fact that you can send the same idea using different patterns in the same language or even different patterns by using another language shows that the ideas themselves are independent from the pattern which is sent across the communication channel. That is how we know that the idea “contained” in the pattern is defined independent of the pattern itself. We could even state the same meaning in a different way – “Do you have the ability to comprehend what these symbols mean?” Either way, the idea contained in the above pattern (question) can be transferred across a communication channel as an independent pattern of letters. This is referred to as functional semantic specificity – where specific groups of patterns which produce semantic/meaningful function are “islands” of specified patterns within a set of all possible patterns.
What about the event: ajdjf9385786gngspaoi-whwtht0wuetghskmnvs-12? Is that pattern specified? Well, is it compressible? Hardly, if at all. Can it be stated any other way, in terms of definition, function, or formulation? Well, this question can only be answered through cryptographic methods. If there is no function, or formulation (description or formulation which is independent of the pattern) then the only way to deliver it across a communication channel is to actually send the pattern itself. Of course, you could send a phonetic spelling of each unit across the communication channel and this would show that the pattern of each unit is specified, but that doesn’t specify the pattern as a whole – the pattern that emerges from the string of specified units. Therefore, until non-arbitrary cryptographic evidence states otherwise, the above pattern is not specified.
To sum up, as has been shown above, a specified pattern is described, independent of the event in question, by the rules of a system. As such, explanations other than chance are to be posited which can create informational patterns that are described by the rules of a system. However, specificity is still not quite good enough by itself to determine previous intelligent cause.
What is needed is a specification, which is a highly improbable specified pattern. But, how do we determine what is highly improbable? We take a look at the available probabilistic resources – that is how many bit operations were needed or used in order to arrive at the specified pattern. Measuring the specified pattern against how long it took to arrive at that pattern and how may different trials where associated with that pattern will tell us if the specified pattern is also highly improbable – beyond all probabilistic resources necessary to generate the specified pattern by chance.
This measure of complexity, which is added onto specificity to create CSI or a specification, is akin to the complexity (improbability) of drawing a royal flush 5 times in a row. Hmmm ... should we begin looking into causes other than chance?
Now, let’s take a look at measuring for a specification. First, it must be understood that this only applies to patterns which can be measured probabilistically. Since a specification includes, but is not limited to function, I will use an example of specification based on compressibility, since compressibility is a way of independently formulating a certain pattern as I have shown above.
Let’s return to our first example – the long pattern of 1s. Dembski has stated that the higher the compressibility of the pattern, the higher the specificity. Why is that the case? That is the case, since the less compressible the pattern is, the more it becomes algorithmically complex and the more random it becomes. These algorithmically complex patterns are the types of patterns that will be generated by random processes.
For example: It is way more likely for a pattern with the same compressibility as “100101111101100001010001011100" to be generated by a random flip of the coin than for a pattern with the same compressibility as “111111111111111111111111111111" to be generated. Why? Because, the only other pattern that can be formulated with the same algorithmic compressibility as the pattern of 30 1s is “000000000000000000000000000000,” which can be compressed to “print 0 X 30.” However, there are many more other patterns with the same compressibility as the first, more random pattern (assortment of 50% 1s and 50% 0s which are sorted in a truly random fashion as per the rules of statistics*). So, it is easier for chance the “find” one of the less compressible patterns, because there are more patterns with the same lower compressibility than there are patterns with the same higher compressibility -- in the above case there are only 2 out of 1,073,741,824 patterns which have the highest compressibility and those two patterns are shown above in the repeating 1s and the repeating 0s.
Now, let’s calculate the pattern: 111111111111111111111111111111 and see if it is a specification.
First we need to calculate it’s specificity. That is done by multiplying its probability (as 1 in 1,073,741,824) with how many other patterns have that same compressibility (in this case only one other pattern as shown above).
So, the specificity of this pattern = 2 * 1/1073741824
Now, in order to move on to finding out if we have a specification here, we must first understand the context in which we actually found this pattern. Let’s say that we are running a search on a string of characters 30 bits long – the size of the pattern above. Let’s say we start at a random point such as: 111100001010011110000110111101. Now, let’s also say that in 30 bit flips (30 operations), we arrive at the above pattern of thirty 1s which we are calculating. Is it reasonable to presume that the pattern was arrived at by pure chance? Let’s make the calculation and find out.
Specification: ?>1 = -log2 [number of bit operations * specificity] = ?
Specification: ?>1 = -log2 [30 * 2/1073741824] = -log2
[0.0000000558793544769287109375]
Specification: ?>1 = approx. 24 bits of CSI = a specification
It is obvious that the discovery of the pattern was not random, but was somehow guided. It can be rejected as being the result of strictly random processes for 2 reasons
1. it is not in the nature of random processes to generate specificity -- in this case regularities (high compressibility) -- and
2. because the number of random bit operations falls extremely short of the probability of arriving at the end pattern taking in account the number of trials (probabilistic resources). To say that this pattern was the result of chance would be to resort to a “chance of the gaps” type of argument. In fact, it has been shown that evolutionary algorithms (which are non-random) are a necessity to arrive at specifications such as the one above. In fact, Dembski and Marks have discussed such algorithms on their evolutionary informatics site.
Now, let’s compare that example to a pattern that has the same probability as a higher percentage of all possible combinations, such as the pattern: 110101000001101001110001111010, which is highly incompressible (thus more algorithmically complex and more random). If you wanted to “compress” this pattern you may end up with: “print 1 X 2, 0101, 0 X 5, 1 X 2, 01, 0 X 2, 1 X 3, 0 X 3, 1 X 4, 010"
Now, let’s get to the calculation.
Specification: ?>1 = -log2 [number of bit flips * number of specified patterns * probability]
Specification: ?>1 = -log2 [30 * X * 1/1,073,741,824]
Now, I haven’t included the number of specified patterns which are as compressible as the above random number, since I am not sure exactly how to calculate that number. However, you could theoretically calculate all other possible compressed patterns which contain the same amount of information and are just as random as the compressed pattern above, and as far as I am aware, that is precisely with what algorithmic information theory deals. Dembski has shown the math involved with algorithmic compressibility in “Specification: The Pattern that Signifies Intelligence” and has shown and concluded: “To sum up, the collection of algorithmically compressible (and therefore nonrandom) sequences has small probability among the totality of sequences, so that observing such a sequence is reason to look for explanations other than chance.”
The corollary to what Dembski has summed up is that algorithmically incompressible sequences make up the rest of the sequence space and thus have a large probability among the totality of sequences. So, we do know that the “X” in the above equation will be a very large number and will produce a less than one amount of CSI and there will be no specification.
For example, even if only 1/32 of all possible patterns are algorithmically random, then the equation would play out as follows:
Specification: ?>1 = -log2 [30 * 33,554,432 * 1/1,073,741,824]
Specification: ?>1 = -log2 [.9375]
Specification: ?>1 = approx. .093 = not > 1 = not a specification
So far, I’ve only shown an example of a specification that was not algorithmically complex. But now, let’s briefly discuss a specification that is algorithmically complex (non-repetitive) and also pseudo-random.
In this case, pseudo-random patterns are those patterns which are algorithmically complex and thus they appear to be random, however, they are specified because of function or they match some independent pattern as set by a system of rules ie: mathematical, linguistic, rules of an information processing system, etc. Basically, as I have stated above, these are the types of patterns which form “islands” of function/pseudo-randomness within a sea of all possible patterns.
When measuring for a functional specification (within a set of functional "islands"), you apply the same equation, however, when measuring the specificity you take into account all other FUNCTIONAL patterns (able to be processed into function *by the system in question*) that have the same probability of appearance as the pattern in question. You do that instead of taking into account all equally probable compressible patterns, since you are now measuring for functional specificity as opposed to compressible specificity. Therefore, you can only measure for functional specificity and then a specification based upon a high understanding of the system and pattern in question.
Furthermore, according to the NFL Theorem, an evolutionary algorithm based on problem specific information is necessary in order to arrive at better than chance performance, which is exactly what a specification is calculating.
The next question: will a random set of laws cause an information processing system and evolutionary algorithm to randomly materialize? According to recent work on Conservation of Information Theorems ID theorists state that the answer is "NO!" In fact, getting consistently better than chance results without previous guiding, problem specific information is to information theory what perpetual motion free energy machines are to physics. To continue to say life was a result of chance would be to appeal to a “chance of the gaps” non-explanation. Physicist Brian Greene states (I found this on God3's Blog [link]):
To invoke multiple universes to explain phenomenon within our universe is merely inflating one’s probabilistic resources beyond reason, thus causing a halt on further investigation since a chance of the gaps “explanation” has already been given.
As a professor Hassofer put it:
So, merely observe an information processing system and evolutionary algorithm self- generate from a truly random set of laws and the foundation of ID Theory is falsified. Or, show how that is even theoretically possible. Science is based on observation and testing hypothesis, and data trumps every time. See the discussion on the Conservation of Information Theorem [link] for why Evolutionary Algorithms will not generate themselves out of a random set of laws.
First, we must begin by defining a specified pattern. What is the basic idea behind a specified pattern? Why must we define a specified pattern? It is because within patterns, there can be something (potentially non-random characteristics such as function) which separates one set of patterns from all other possible patterns.
Dembski begins to discuss a specified pattern by stating: “The crucial difference between (R) [a random pattern] and (ψR) [a pseudo-random pattern -- the Champernowne sequence] is that (ψR) exhibits a simple, easily described pattern whereas (R) does not. To describe (ψR), it is enough to note that this sequence lists binary numbers in increasing order. By contrast, (R) cannot, so far as we can tell, be described any more simply than by repeating the sequence.” I will continue to discuss the difference between random and pseudo random patterns (as it relates to specificity) further on.
Dr. Dembski describes specified patterns as those patterns which can be described and formulated independent of the event (pattern) in question.
Stephen Myers expounds upon this subject in his paper, “DNA and the Origin of Life: ...”:
“Moreover, given their [proteins’] irregularity, it seemed unlikely that a general chemical law or regularity could explain their assembly. Instead, as Jacques Monod has recalled, molecular biologists began to look for some source of information or “specificity” within the cell that could direct the construction of such highly specific and complex structures. To explain the presence of the specificity and complexity in the protein, as Monod would later insist, “you absolutely needed a code.”21"
... and ...
“In essence, therefore, Shannon’s theory remains silent on the important question of whether a sequence of symbols is functionally specific or meaningful. Nevertheless, in its application to molecular biology, Shannon information theory did succeed in rendering rough quantitative measures of the information-carrying capacity or syntactic information (where those terms correspond to measures of brute complexity).33 ... In essence, therefore, Shannon’s theory remains silent on the important question of whether a sequence of symbols is functionally specific or meaningful. Nevertheless, in its application to molecular biology, Shannon information theory did succeed in rendering rough quantitative measures of the information-carrying capacity or syntactic information (where those terms correspond to measures of brute complexity).33"
... and ...
“Since the late 1950s, biologists have equated the “precise determination of sequence” with the extra-information-theoretic property of specificity or specification. Biologists have defined specificity tacitly as “necessary to achieve or maintain function.” They have determined that DNA base sequences, for example, are specified not by applying information theory but by making experimental assessments of the function of those sequences within the overall apparatus of gene expression.36 Similar experimental considerations established the functional specificity of proteins. Further, developments in complexity theory have now made possible a fully general theoretical account of specification, one that applies readily to biological systems. In particular, recent work by mathematician William Dembski has employed the notion of a rejection region from statistics to provide a formal complexity-theoretic account of specification. According to Dembski, a specification occurs when an event or object (a) falls within an independently given pattern or domain, (b) “matches” or exemplifies a conditionally independent pattern, or (c) meets a conditionally independent set of functional requirements.37"
This ability to formulate a pattern independent of the event in question can be understood with a couple of examples. But, we must first start with a communication system. Why? So that we can exchange a pattern across that communication system and see if it can indeed be formulated independent of the pattern in question.
Let’s say that we wanted to send the pattern: 111111111111111111111111111111 across our communication system. First, can we compress it? Yes we can. How? By defining or formulating it differently. On way to do that is by compressing the pattern to “print 1 X 30.” Thus we can send this information across the communication channel and the receiving party will receive the exact same information than if we had sent the original pattern of 1s. Therefore, the original pattern can be defined independently of the pattern itself. So, compressibility is one way to define a specified pattern.
Now, let’s look at another pattern: 1 2 3 5 7 11 13 17 19 23 ... . Is this a specified pattern? In order to evaluate this and the next examples of patterns, we need to understand the difference between what Dembski calls a pseudo-random and a random pattern. A random pattern is algorithmically complex and can not be defined or formulated according to a system of rules whereas a pseudo-random pattern merely looks random, since it is algorithmically complex (not regular as the last example of 1s), however it is not random since it can be defined and formulated as a separate pattern according to a system of rules. This usually entails meaning or function.
So, let’s return to our example of the pattern: 1 2 3 5 7 11 13 17 19 23 ... . Can we define or formulate it in a different way yet retain the same information? Sure we can. This pattern conforms to a system of mathematical rules and can be defined or formulated differently than by merely repeating the pattern. We can send: “all numbers, beginning with one, which are divisible only by themselves and one” or its equivalent mathematical notation across the communication channel and the other end would be able to reconstruct the original pattern -- a sequence of prime numbers.
Now, let’s take a look at proteins. When it comes to measuring specificity, this is exactly like measuring specificity in a meaningful sentence, as I will soon show. Functional specificity merely separates functional pattern “islands” from the sea of random possible patterns. When specific proteins are brought together, you can have a pattern which creates function [http://en.wikipedia.org/wiki/Function_%28biology%29]. [link] That functional pattern itself is formulated by information contained in DNA which is encoded into RNA and decoded into the specific system of functional proteins. The functional pattern as the event in question is defined independently as a pattern of nucleic acids. Thus, the independent formulation for the system of protein function is sent across a communication channel as RNA which is the independent formulation of the function (I have provided a definition of “function” [here “philosophical foudations”] [link]). The RNA is independent of the function itself for which it codes (that is, the information for formulating the function doesn’t come from the protein pattern itself as per Central Dogma of Biology [http://en.wikipedia.org/wiki/Central_dogma_of_molecular_biology] [link]). You don’t send the function itself across the communication channel, you send an informational pattern. Again, that is what is referred to as functional specificity.
What about the pattern: “Can you understand this?” This pattern is most definitely specified, since it can be defined according to a pre-set English dictionary and produce a function through meaning/ideas. How do you pass an idea/meaning through a communication channel? You do so, by using informational units derived from a pre-set system which are separate from the actual pattern of ideas/meaning and which can be processed according to the rules of that pre-set system at the other end of the communication channel. The fact that you can send the same idea using different patterns in the same language or even different patterns by using another language shows that the ideas themselves are independent from the pattern which is sent across the communication channel. That is how we know that the idea “contained” in the pattern is defined independent of the pattern itself. We could even state the same meaning in a different way – “Do you have the ability to comprehend what these symbols mean?” Either way, the idea contained in the above pattern (question) can be transferred across a communication channel as an independent pattern of letters. This is referred to as functional semantic specificity – where specific groups of patterns which produce semantic/meaningful function are “islands” of specified patterns within a set of all possible patterns.
What about the event: ajdjf9385786gngspaoi-whwtht0wuetghskmnvs-12? Is that pattern specified? Well, is it compressible? Hardly, if at all. Can it be stated any other way, in terms of definition, function, or formulation? Well, this question can only be answered through cryptographic methods. If there is no function, or formulation (description or formulation which is independent of the pattern) then the only way to deliver it across a communication channel is to actually send the pattern itself. Of course, you could send a phonetic spelling of each unit across the communication channel and this would show that the pattern of each unit is specified, but that doesn’t specify the pattern as a whole – the pattern that emerges from the string of specified units. Therefore, until non-arbitrary cryptographic evidence states otherwise, the above pattern is not specified.
To sum up, as has been shown above, a specified pattern is described, independent of the event in question, by the rules of a system. As such, explanations other than chance are to be posited which can create informational patterns that are described by the rules of a system. However, specificity is still not quite good enough by itself to determine previous intelligent cause.
What is needed is a specification, which is a highly improbable specified pattern. But, how do we determine what is highly improbable? We take a look at the available probabilistic resources – that is how many bit operations were needed or used in order to arrive at the specified pattern. Measuring the specified pattern against how long it took to arrive at that pattern and how may different trials where associated with that pattern will tell us if the specified pattern is also highly improbable – beyond all probabilistic resources necessary to generate the specified pattern by chance.
This measure of complexity, which is added onto specificity to create CSI or a specification, is akin to the complexity (improbability) of drawing a royal flush 5 times in a row. Hmmm ... should we begin looking into causes other than chance?
Now, let’s take a look at measuring for a specification. First, it must be understood that this only applies to patterns which can be measured probabilistically. Since a specification includes, but is not limited to function, I will use an example of specification based on compressibility, since compressibility is a way of independently formulating a certain pattern as I have shown above.
Let’s return to our first example – the long pattern of 1s. Dembski has stated that the higher the compressibility of the pattern, the higher the specificity. Why is that the case? That is the case, since the less compressible the pattern is, the more it becomes algorithmically complex and the more random it becomes. These algorithmically complex patterns are the types of patterns that will be generated by random processes.
For example: It is way more likely for a pattern with the same compressibility as “100101111101100001010001011100" to be generated by a random flip of the coin than for a pattern with the same compressibility as “111111111111111111111111111111" to be generated. Why? Because, the only other pattern that can be formulated with the same algorithmic compressibility as the pattern of 30 1s is “000000000000000000000000000000,” which can be compressed to “print 0 X 30.” However, there are many more other patterns with the same compressibility as the first, more random pattern (assortment of 50% 1s and 50% 0s which are sorted in a truly random fashion as per the rules of statistics*). So, it is easier for chance the “find” one of the less compressible patterns, because there are more patterns with the same lower compressibility than there are patterns with the same higher compressibility -- in the above case there are only 2 out of 1,073,741,824 patterns which have the highest compressibility and those two patterns are shown above in the repeating 1s and the repeating 0s.
Now, let’s calculate the pattern: 111111111111111111111111111111 and see if it is a specification.
First we need to calculate it’s specificity. That is done by multiplying its probability (as 1 in 1,073,741,824) with how many other patterns have that same compressibility (in this case only one other pattern as shown above).
So, the specificity of this pattern = 2 * 1/1073741824
Now, in order to move on to finding out if we have a specification here, we must first understand the context in which we actually found this pattern. Let’s say that we are running a search on a string of characters 30 bits long – the size of the pattern above. Let’s say we start at a random point such as: 111100001010011110000110111101. Now, let’s also say that in 30 bit flips (30 operations), we arrive at the above pattern of thirty 1s which we are calculating. Is it reasonable to presume that the pattern was arrived at by pure chance? Let’s make the calculation and find out.
Specification: ?>1 = -log2 [number of bit operations * specificity] = ?
Specification: ?>1 = -log2 [30 * 2/1073741824] = -log2
[0.0000000558793544769287109375]
Specification: ?>1 = approx. 24 bits of CSI = a specification
It is obvious that the discovery of the pattern was not random, but was somehow guided. It can be rejected as being the result of strictly random processes for 2 reasons
1. it is not in the nature of random processes to generate specificity -- in this case regularities (high compressibility) -- and
2. because the number of random bit operations falls extremely short of the probability of arriving at the end pattern taking in account the number of trials (probabilistic resources). To say that this pattern was the result of chance would be to resort to a “chance of the gaps” type of argument. In fact, it has been shown that evolutionary algorithms (which are non-random) are a necessity to arrive at specifications such as the one above. In fact, Dembski and Marks have discussed such algorithms on their evolutionary informatics site.
Now, let’s compare that example to a pattern that has the same probability as a higher percentage of all possible combinations, such as the pattern: 110101000001101001110001111010, which is highly incompressible (thus more algorithmically complex and more random). If you wanted to “compress” this pattern you may end up with: “print 1 X 2, 0101, 0 X 5, 1 X 2, 01, 0 X 2, 1 X 3, 0 X 3, 1 X 4, 010"
Now, let’s get to the calculation.
Specification: ?>1 = -log2 [number of bit flips * number of specified patterns * probability]
Specification: ?>1 = -log2 [30 * X * 1/1,073,741,824]
Now, I haven’t included the number of specified patterns which are as compressible as the above random number, since I am not sure exactly how to calculate that number. However, you could theoretically calculate all other possible compressed patterns which contain the same amount of information and are just as random as the compressed pattern above, and as far as I am aware, that is precisely with what algorithmic information theory deals. Dembski has shown the math involved with algorithmic compressibility in “Specification: The Pattern that Signifies Intelligence” and has shown and concluded: “To sum up, the collection of algorithmically compressible (and therefore nonrandom) sequences has small probability among the totality of sequences, so that observing such a sequence is reason to look for explanations other than chance.”
The corollary to what Dembski has summed up is that algorithmically incompressible sequences make up the rest of the sequence space and thus have a large probability among the totality of sequences. So, we do know that the “X” in the above equation will be a very large number and will produce a less than one amount of CSI and there will be no specification.
For example, even if only 1/32 of all possible patterns are algorithmically random, then the equation would play out as follows:
Specification: ?>1 = -log2 [30 * 33,554,432 * 1/1,073,741,824]
Specification: ?>1 = -log2 [.9375]
Specification: ?>1 = approx. .093 = not > 1 = not a specification
So far, I’ve only shown an example of a specification that was not algorithmically complex. But now, let’s briefly discuss a specification that is algorithmically complex (non-repetitive) and also pseudo-random.
In this case, pseudo-random patterns are those patterns which are algorithmically complex and thus they appear to be random, however, they are specified because of function or they match some independent pattern as set by a system of rules ie: mathematical, linguistic, rules of an information processing system, etc. Basically, as I have stated above, these are the types of patterns which form “islands” of function/pseudo-randomness within a sea of all possible patterns.
When measuring for a functional specification (within a set of functional "islands"), you apply the same equation, however, when measuring the specificity you take into account all other FUNCTIONAL patterns (able to be processed into function *by the system in question*) that have the same probability of appearance as the pattern in question. You do that instead of taking into account all equally probable compressible patterns, since you are now measuring for functional specificity as opposed to compressible specificity. Therefore, you can only measure for functional specificity and then a specification based upon a high understanding of the system and pattern in question.
Furthermore, according to the NFL Theorem, an evolutionary algorithm based on problem specific information is necessary in order to arrive at better than chance performance, which is exactly what a specification is calculating.
The next question: will a random set of laws cause an information processing system and evolutionary algorithm to randomly materialize? According to recent work on Conservation of Information Theorems ID theorists state that the answer is "NO!" In fact, getting consistently better than chance results without previous guiding, problem specific information is to information theory what perpetual motion free energy machines are to physics. To continue to say life was a result of chance would be to appeal to a “chance of the gaps” non-explanation. Physicist Brian Greene states (I found this on God3's Blog [link]):
‘If true, the idea of a multiverse would be a Copernican Revolution realized on a cosmic scale. It would be a rich and astounding upheaval, but one with potentially hazardous consequences. Beyond the inherent difficulty in assessing its validity, when should we allow the multiverse framework to be invoked in lieu of a more traditional scientific explanation? Had this idea surfaced a hundred years ago, might researchers have chalked up various mysteries to how things just happen to be in our corner of the multiverse and not pressed on to discover all the wondrous science of the last century? …The danger, if the multiverse idea takes root, is that researchers may too quickly give up the search for underlying explanations. When faced with seemingly inexplicable observations, researchers may invoke the framework of the multiverse prematurely – proclaiming some phenomenon or other to merely reflect conditions in our own bubble universe and thereby failing to discover the deeper understanding that awaits us. ‘
To invoke multiple universes to explain phenomenon within our universe is merely inflating one’s probabilistic resources beyond reason, thus causing a halt on further investigation since a chance of the gaps “explanation” has already been given.
As a professor Hassofer put it:
“The problem [of falsifiability of a probabilistic statement] has been dealt with in a recent book by G. Matheron, entitled Estimating and Choosing: An Essay on Probability in Practice (Springer-Verlag, 1989). He proposes that a probabilistic model be considered falsifiable if some of its consequences have zero (or in practice very low) probability. If one of these consequences is observed, the model is then rejected.
‘The fatal weakness of the monkey argument, which calculates probabilities of events “somewhere, sometime”, is that all events, no matter how unlikely they are, have probability one as long as they are logically possible, so that the suggested model can never be falsified. Accepting the validity of Huxley’s reasoning puts the whole probability theory outside the realm of verifiable science. In particular, it vitiates the whole of quantum theory and statistical mechanics, including thermodynamics, and therefore destroys the foundations of all modern science. For example, as Bertrand Russell once pointed out, if we put a kettle on a fire and the water in the kettle froze, we should argue, following Huxley, that a very unlikely event of statistical mechanics occurred, as it should “somewhere, sometime”, rather than trying to find out what went wrong with the experiment!’”
So, merely observe an information processing system and evolutionary algorithm self- generate from a truly random set of laws and the foundation of ID Theory is falsified. Or, show how that is even theoretically possible. Science is based on observation and testing hypothesis, and data trumps every time. See the discussion on the Conservation of Information Theorem [link] for why Evolutionary Algorithms will not generate themselves out of a random set of laws.
Sunday, February 10, 2008
Intelligence Law and Chance Working Together
This post is part II of the Philosophical Foundation for ID Theory.
Someone posed the following questions to me:
Those questions bear the same weight as the following question: “Will a random set of laws (that is, only law and chance) generate information processing systems, CSI, evolutionary algorithms, systems based on engineering principles (control nature without being defined by law), and intelligence?” Some theorists state that based on the NFL Theorems and COI Theorems, the answer is a resounding “NO!” Thus, we need another known causal phenomenon which has been observed creating those types of systems and our investigation into ID Theory continues.
You don’t need to formally describe law and chance, unless by formally describe you mean formally analyse how to measure law and chance. Either way, all you need to do is formally describe CSI and show that neither chance (measured and described as randomness) nor law (measured and described as regularities) can characterize this event (which can be made as a potentially falsifiable hypothesis). Once it is shown that intelligence does produce algorithmically complex CSI, and it is shown that intelligence itself is necessarily founded upon this information, then you have a closed loop from intelligence to information to intelligence, with no room for *only* law and chance. Thus, you must tentatively include intelligence as a third mode of explanation until the above hypothesis -- that a random set of laws (merely law and chance) will not generate information processing systems, etc. and intelligence is a necessary cause of those systems -- is falsified. In fact, we see many systems every day which we know are not and most probably could not have been created by mere chance and law not being “filtered” through intelligence as previously defined. Thus, such systems do exist.
Furthermore, you don’t have to completely negate law and chance operating within a system in order to arrive at the conclusion that the system itself is intelligently designed. In fact, that is not how reality works. Even a car, although it is intelligently designed, can show the effects of law and chance. Look at a rusty old beater car. You can’t explain the “rust” feature by design, but that doesn’t negate the fact that the car itself was intelligently designed. Furthermore, you can’t point to the rust as a “faulty feature” and thus arrive at the conclusion that the car could have or must have been created by only chance and law.
Another example is if random processes create a new function or strengthen an older function by damaging or clogging a system, does that prove that the original system was the result of only chance and law? Let’s look at a lock and key mechanism. Just because natural processes can increase the security function of an ancient castle door by “gunking” up the lock, making it unopenable even with the original key, does that mean that unguided law and chance can create the castle, the door, the lock, and the key? Of course not!
I have a further question for you. How can you show that chance and law are possible apart from an intelligently designed system? It seems that you need an information processing system to generate law and you must have law first, to generate chance. Are information processing systems possible without intelligent programming? Just some food for thought. I personally don’t think that you can define any one causal phenomenon at the pure exclusion of the others.
In fact, we see intelligence, law, and chance working together the best within evolutionary algorithms, since intelligence inputs the information necessary to arrive at optimization problems efficiently by programming the laws to make use of controlled chance and do all the dirty work for him/her.
Of course, intelligence needs to be defined, and that may be the hardest part, yet not intractable. I personally define intelligence as “a system which can plan into the future and then sufficiently engineer a solution to accomplish a goal.”
Intelligence can basically be summed up in the ability to produce a goal oriented procedure by utilizing (sufficiently organizing/programming) stochastic (law and chance) processes. We know that we as humans can do this and so intelligence does exist as a causal entity alongside chance and law.
Furthermore, if it is true that the NFLT and Conservation of Information Theorem both show the necessity of prior problem specific information (attained by neither law nor chance) in order to produce CSI, then it can be argued that active information is the formal measurement of that (problem specific information) which is neither law nor chance. In this case, active information is a measurement of one aspect of intelligence in a specific situation (efficient optimization algorithms).
The above stands until anyone else can show that problem specific information (and thus evolutionary algorithms) necessary to generate intelligent systems can be generated by a random assortment of laws – that is laws which are not organized by previous intelligence -- pure law and chance. Until you show that intelligence is not necessarily the result of intelligence, it is quite obvious based on observation, that intelligence is indeed a causal phenomenon alongside chance and law.
Sum up:
1. Intelligence as a foresight using system does exist. The ability to organize law and chance for a future determined goal does exist. Just to be clear, this has nothing to do with “free will.” This only concerns the ability to plan into the future. Whether this ability is a free choice or not is a completely different issue than the fact that this ability exists.
2. There does exist systems which are only generated if “run through” intelligence. Intelligence does generate system where forethought is a necessity.
3. CSI has been formally described such that, in accordance with NFL Theorems, it is for all practical purposes not attainable through only chance and law (random set of laws), just as perpetual motion free energy machines or the past hour of the universe running in reverse are practically not attainable based on our understanding of cause and effect and the flow of energy and information. This has been put foreward as a hypothesis that intelligence is a necessary cause of CSI, since CSI has been observed being created by intelligence. This is obviously potentially falsifiable.
4. Therefore, Intelligence must be tentatively included alongside chance and law until point 3 is falsified.
Someone posed the following questions to me:
“"Dembski’s characterization of design as a third mode of explanation apart from chance and/or law is one of the most fundamental problems with his approach. How do you formally (i.e. mathematically) describe chance and law such that their disjunction doesn’t characterize all conceivable events? How do you show that design, or intelligent agency as you say, isn’t an instance of chance and/or law? More importantly, how do you show that it could possibly not be an instance of chance and/or law?"
Those questions bear the same weight as the following question: “Will a random set of laws (that is, only law and chance) generate information processing systems, CSI, evolutionary algorithms, systems based on engineering principles (control nature without being defined by law), and intelligence?” Some theorists state that based on the NFL Theorems and COI Theorems, the answer is a resounding “NO!” Thus, we need another known causal phenomenon which has been observed creating those types of systems and our investigation into ID Theory continues.
You don’t need to formally describe law and chance, unless by formally describe you mean formally analyse how to measure law and chance. Either way, all you need to do is formally describe CSI and show that neither chance (measured and described as randomness) nor law (measured and described as regularities) can characterize this event (which can be made as a potentially falsifiable hypothesis). Once it is shown that intelligence does produce algorithmically complex CSI, and it is shown that intelligence itself is necessarily founded upon this information, then you have a closed loop from intelligence to information to intelligence, with no room for *only* law and chance. Thus, you must tentatively include intelligence as a third mode of explanation until the above hypothesis -- that a random set of laws (merely law and chance) will not generate information processing systems, etc. and intelligence is a necessary cause of those systems -- is falsified. In fact, we see many systems every day which we know are not and most probably could not have been created by mere chance and law not being “filtered” through intelligence as previously defined. Thus, such systems do exist.
Furthermore, you don’t have to completely negate law and chance operating within a system in order to arrive at the conclusion that the system itself is intelligently designed. In fact, that is not how reality works. Even a car, although it is intelligently designed, can show the effects of law and chance. Look at a rusty old beater car. You can’t explain the “rust” feature by design, but that doesn’t negate the fact that the car itself was intelligently designed. Furthermore, you can’t point to the rust as a “faulty feature” and thus arrive at the conclusion that the car could have or must have been created by only chance and law.
Another example is if random processes create a new function or strengthen an older function by damaging or clogging a system, does that prove that the original system was the result of only chance and law? Let’s look at a lock and key mechanism. Just because natural processes can increase the security function of an ancient castle door by “gunking” up the lock, making it unopenable even with the original key, does that mean that unguided law and chance can create the castle, the door, the lock, and the key? Of course not!
I have a further question for you. How can you show that chance and law are possible apart from an intelligently designed system? It seems that you need an information processing system to generate law and you must have law first, to generate chance. Are information processing systems possible without intelligent programming? Just some food for thought. I personally don’t think that you can define any one causal phenomenon at the pure exclusion of the others.
In fact, we see intelligence, law, and chance working together the best within evolutionary algorithms, since intelligence inputs the information necessary to arrive at optimization problems efficiently by programming the laws to make use of controlled chance and do all the dirty work for him/her.
Of course, intelligence needs to be defined, and that may be the hardest part, yet not intractable. I personally define intelligence as “a system which can plan into the future and then sufficiently engineer a solution to accomplish a goal.”
Intelligence can basically be summed up in the ability to produce a goal oriented procedure by utilizing (sufficiently organizing/programming) stochastic (law and chance) processes. We know that we as humans can do this and so intelligence does exist as a causal entity alongside chance and law.
Furthermore, if it is true that the NFLT and Conservation of Information Theorem both show the necessity of prior problem specific information (attained by neither law nor chance) in order to produce CSI, then it can be argued that active information is the formal measurement of that (problem specific information) which is neither law nor chance. In this case, active information is a measurement of one aspect of intelligence in a specific situation (efficient optimization algorithms).
The above stands until anyone else can show that problem specific information (and thus evolutionary algorithms) necessary to generate intelligent systems can be generated by a random assortment of laws – that is laws which are not organized by previous intelligence -- pure law and chance. Until you show that intelligence is not necessarily the result of intelligence, it is quite obvious based on observation, that intelligence is indeed a causal phenomenon alongside chance and law.
Sum up:
1. Intelligence as a foresight using system does exist. The ability to organize law and chance for a future determined goal does exist. Just to be clear, this has nothing to do with “free will.” This only concerns the ability to plan into the future. Whether this ability is a free choice or not is a completely different issue than the fact that this ability exists.
2. There does exist systems which are only generated if “run through” intelligence. Intelligence does generate system where forethought is a necessity.
3. CSI has been formally described such that, in accordance with NFL Theorems, it is for all practical purposes not attainable through only chance and law (random set of laws), just as perpetual motion free energy machines or the past hour of the universe running in reverse are practically not attainable based on our understanding of cause and effect and the flow of energy and information. This has been put foreward as a hypothesis that intelligence is a necessary cause of CSI, since CSI has been observed being created by intelligence. This is obviously potentially falsifiable.
4. Therefore, Intelligence must be tentatively included alongside chance and law until point 3 is falsified.
Science Deals with Evidence
I’ve heard some people make the outrageous claim that there isn’t any evidence of intelligent cause for life. Let’s examine this, and see if these people are merely being selective hyper- sceptics or if they really do have a point that there is no evidence of intelligent cause within life.
First let’s briefly look at intelligence. What is intelligence? There are many ways to discuss intelligence and many different properties that can be attributed to intelligence. However, as far as I have seen, the two most common “tests” of intelligence is the ability to learn and the ability to use functional information and knowledge.
What is learning? Well, one thing is true about learning ... that is, when you learn something, the useful information and knowledge that you possess increases. If an AI system learns, then it increase its useable information as a functional response to its environment. One thing that distinguishes intelligence from other systems is its ability to increase its functional information
and knowledge content as it interacts with its environment.
Intelligence can then use this information to manipulate its environment. That seems to be what an IQ test actually looks for ... the ability to apply information to situations and problems. If intelligence can apply information to environmental situations, it will then be able to utilize stochastic processes (natural law and chance) to produce a goal oriented procedure to solve a particular problem or cause a pre-planned result which is not definable by natural theoretical law and not reasonably attributed to pure chance.
Therefore, ID Theory postulates that certain effects of intelligence can be reliably separated from theoretical law and chance – intelligent cause is detectable. Since intelligence can basically be summed up in the ability to produce a goal oriented procedure by utilizing (sufficiently organizing/programming) stochastic processes, intelligence can be equated with any type of teleological process – teleology being defined as: “the fact or character attributed to nature or natural processes of being directed toward an end or shaped by a purpose.” Neither chance nor a random assortment of law are teleological. However, intelligence is marked by the ability to plan into the future and then sufficiently engineer a solution to accomplish a goal. So the debate concerning Intelligent Design Theory is fundamentally one of accident verses teleology – that is, chance assemblage of law verses intelligent assemblage and fine tuning of law (guidance toward a pre-planned goal). And both of these positions are testable by experimenting with information processing systems.
ID Theory hypothesizes that the systems below are reliable indicators of previous teleological/intelligent cause because they are not theoretically definable by law, neither reasonably attributed nor observed to have arisen by pure chance, yet are (as a positive) observed to have been caused by intelligent systems using foresight to accomplish future goals:
-CSI (Specifications) [link]
-Communication system, Information processor, Coded Information, Coding/Decoding System (Definition of a code: Given a source with probability space [Omega, A, p(A)] and a receiver with probability space [Omega, B, p(B)], then a unique mapping of the letters of alphabet A onto letters of alphabet B is called a code. Here p(A) is the probability vector of the elements of alphabet A and p (B) is the probability vector of the elements of alphabet B. (Perlwitz, Burks and Waterman, 1988). (According to this definition, which is very simple, DNA is a code.)
-Evolutionary algorithms [link], which must be guided by problem-specific active information (knowledge of the problem/targets programmed into the behaviour of the algorithm).
... as well ...
Convergent evolution [link][link]is the observation of many separate biological evolutionary trials converging upon the same forms and functions. Convergent evolution provides evidence that the evolutionary process is focussed upon non-accidental end functions (targets) and is consistent with the hypothesis that there is an ultimate evolutionary end point (the Omega Point) -- the ultimate target of evolution. Convergent evolution shows the constraint of evolution, after multiple separate runs, to the same targets of highly improbable form and function. This provides evidence that biological evolution follows what has already been discovered with the NFL Theorems – that evolutionary algorithms only work if the search procedure is guided to pre- set problems by problem specific information. It’s one thing for evolution to find a highly improbable form or function once, however, when it finds it on multiple separate trials we have evidence that it is indeed a process which is guided toward a palate of pre-determined functions which are then sorted by the environment (natural selection).
As an extra, here is Hameroff’s blog [link] discussing his testable and fasifiable model/hypothesis re: a potential foundation for a quantum based intelligent design – that is: consciousness may actually fundamentally be a property of the operation of space-time’s quantum effects. Here’s a good video lecture [google video] [link] by Hameroff giving the basic explanation of his hypothesis/model.
First let’s briefly look at intelligence. What is intelligence? There are many ways to discuss intelligence and many different properties that can be attributed to intelligence. However, as far as I have seen, the two most common “tests” of intelligence is the ability to learn and the ability to use functional information and knowledge.
What is learning? Well, one thing is true about learning ... that is, when you learn something, the useful information and knowledge that you possess increases. If an AI system learns, then it increase its useable information as a functional response to its environment. One thing that distinguishes intelligence from other systems is its ability to increase its functional information
and knowledge content as it interacts with its environment.
Intelligence can then use this information to manipulate its environment. That seems to be what an IQ test actually looks for ... the ability to apply information to situations and problems. If intelligence can apply information to environmental situations, it will then be able to utilize stochastic processes (natural law and chance) to produce a goal oriented procedure to solve a particular problem or cause a pre-planned result which is not definable by natural theoretical law and not reasonably attributed to pure chance.
Therefore, ID Theory postulates that certain effects of intelligence can be reliably separated from theoretical law and chance – intelligent cause is detectable. Since intelligence can basically be summed up in the ability to produce a goal oriented procedure by utilizing (sufficiently organizing/programming) stochastic processes, intelligence can be equated with any type of teleological process – teleology being defined as: “the fact or character attributed to nature or natural processes of being directed toward an end or shaped by a purpose.” Neither chance nor a random assortment of law are teleological. However, intelligence is marked by the ability to plan into the future and then sufficiently engineer a solution to accomplish a goal. So the debate concerning Intelligent Design Theory is fundamentally one of accident verses teleology – that is, chance assemblage of law verses intelligent assemblage and fine tuning of law (guidance toward a pre-planned goal). And both of these positions are testable by experimenting with information processing systems.
ID Theory hypothesizes that the systems below are reliable indicators of previous teleological/intelligent cause because they are not theoretically definable by law, neither reasonably attributed nor observed to have arisen by pure chance, yet are (as a positive) observed to have been caused by intelligent systems using foresight to accomplish future goals:
-CSI (Specifications) [link]
-Communication system, Information processor, Coded Information, Coding/Decoding System (Definition of a code: Given a source with probability space [Omega, A, p(A)] and a receiver with probability space [Omega, B, p(B)], then a unique mapping of the letters of alphabet A onto letters of alphabet B is called a code. Here p(A) is the probability vector of the elements of alphabet A and p (B) is the probability vector of the elements of alphabet B. (Perlwitz, Burks and Waterman, 1988). (According to this definition, which is very simple, DNA is a code.)
-Evolutionary algorithms [link], which must be guided by problem-specific active information (knowledge of the problem/targets programmed into the behaviour of the algorithm).
... as well ...
Convergent evolution [link][link]is the observation of many separate biological evolutionary trials converging upon the same forms and functions. Convergent evolution provides evidence that the evolutionary process is focussed upon non-accidental end functions (targets) and is consistent with the hypothesis that there is an ultimate evolutionary end point (the Omega Point) -- the ultimate target of evolution. Convergent evolution shows the constraint of evolution, after multiple separate runs, to the same targets of highly improbable form and function. This provides evidence that biological evolution follows what has already been discovered with the NFL Theorems – that evolutionary algorithms only work if the search procedure is guided to pre- set problems by problem specific information. It’s one thing for evolution to find a highly improbable form or function once, however, when it finds it on multiple separate trials we have evidence that it is indeed a process which is guided toward a palate of pre-determined functions which are then sorted by the environment (natural selection).
As an extra, here is Hameroff’s blog [link] discussing his testable and fasifiable model/hypothesis re: a potential foundation for a quantum based intelligent design – that is: consciousness may actually fundamentally be a property of the operation of space-time’s quantum effects. Here’s a good video lecture [google video] [link] by Hameroff giving the basic explanation of his hypothesis/model.
Philosophical Foundations of ID to Conservation of Information
My understanding of the foundation of ID Theory. This could also be looked at as the philosophy of ID (upon which the SCIENCE OF ID [link] is founded).
Disclaimer: This post is not philosophical in its entirety, as it does borrow from information theorems and scientific experiments with information and programming (evolutionary algorithms). However, this post and Part II [link] do argue for the basic philosophical foundation necessary to be able to point to intelligence, being independent of the physical laws which arise from or govern matter, as a cause when investigating certain phenomenon.
1. Highly improbable (far beyond UPB), algorithmically complex, specified, coded information (ie. sequential arrangement of nucleotides in RNA) is not caused by the physical properties of the materials (the nucleotides) which are merely used to *transfer* information from DNA to Proteins. IOW, it is not the physical attractive properties of the nucleotides in DNA or RNA or the letters on a page which causes their sequential arrangement. Therefore, at the least, this information is not caused by and transcends the physical properties of matter. This information is caused by a non physical-chemical law.
Furthermore, certain systems exist which follow engineering control principles which are not defined by laws of physics or chemistry, yet control or bound laws and chance to produce function:
“A shaping of boundaries may be said to go beyond a mere fixing of boundaries and establishes a ‘controlling principle.’ It achieves control of the boundaries by imprinting a significant pattern on the boundaries of the system. Or, to use information language, we may say that it puts the system under the control of a non-physical-chemical principle by a profoundly informative intervention.”
--Michael Polanyi, “Life Transcending Physics and Chemistry,” Chemical & Engineering News (21 August 1967): 64.
“In the face of the universal tendency for order to be lost, the complex organization of the living organism can be maintained only if work – involving the expenditure of energy – is performed to conserve the order. The organism is constantly adjusting, repairing, replacing, and this requires energy. But the preservation of the complex, improbable organization of the living creature needs more than energy for the work. It calls for information or instructions on how the energy should be expended to maintain the improbable organization. The idea of information necessary for the maintenance and, as we shall see, creation of living systems is of great utility in approaching the biological problems of reproduction.”
George Gaylord Simpson and William S. Beck, Life: An Introduction to Biology, 2nd ed. (London: Routledge and Kegan, 1965), 145
At the top of this post is Hubert Yockey’s diagram published in the “Journal of Theoretical Biology” showing how life actually follows engineering principles and makes use of an actual communication channel and information processor.
From here on, I am using the term “information” to describe low thermodynamic entropy/high improbability systems, which are not defined by laws of physics or chemistry, that cause functional or semantic specificity.
Function can simply be referred to as the transfer of energy to perform work as a by-product of organized units as per the above quote by George Gaylord Simpson. In man-made structures, function defines the purpose for which that structure was created and in biology function defines that which aids in reproductive and survival success as a result of natural selection.
And, no, this is not a circular definition, since organized units are not necessarily information. IE: snowflakes are highly organized units however they are defined by laws of physics and chemistry and do not cause any functional or semantic specificity.[link]
2. If information is precluded by intelligence (as has been observed), and if intelligence is necessarily founded upon CSI and these systems which follow engineering principles (so far this is the case), then …
3. We have a closed loop traveling from information (not caused by properties of material as per #1) to intelligence (which is then subsequently also not caused *only* by material properties) and back to information, with no room for strictly material causation. Material is only used as a conduit to transfer information.
4. Thus, both information and intelligence may be non-material or at the very least must contain or be caused by a non-material property or law (good bye philosophical materialism). After all, what we observe is that information and intelligence flows through material yet is fundamentally not caused by the properties (physical or chemical laws) of that material.
“But,” you may ask, “when the material is removed, doesn’t the information cease to exist? Isn’t information necessarily dependant on the material which ‘contains’ it? What happens to the information held by the 99% of now extinct species? Does the information still exist?”
Well, it does seem upon first glance that information is necessarily linked to material. Ie: destroy the material conduit and information is destroyed. However, this point does not negate the observation that physical properties of material do not cause information (ie: in biology), which is the main point. Since intelligence does cause information and since intelligence is founded upon information, there must exist some aspect of at least one of these two properties which is not material.
And now a new thought jumps across my mind. Does the destruction of the conduit actually destroy the information? If material only acts as a conduit, and intelligence is actually the source/cause of the information, this means that the information existed within some intelligence or non-material information transferring law before it was sent through the conduit. Destruction of the conduit (material) would only seem to destroy that specific *transfer* of the information.
For example, if I write a coded note which reads “meet at bridge at 0900,” have it delivered to an undercover agent, and upon reading it the agent then burns the note. Has the information been destroyed? No, it most definitely has not been destroyed. The information has actually been transferred and stored within another intelligence. However, let’s say that something went wrong and upon arriving at the bridge, the undercover agent is gunned down. Now, is the information destroyed? No, since I (as the intelligent source) still possess all the necessary information to create the exact same message. The information existed within intelligence before it was transmitted and it continues to exist in the source intelligence after transmission and destruction of the pen and paper note – the transferring medium (material).
In the case of biology, if life and evolution are a necessary result of and are guided by the laws of nature, then wouldn’t the information at the foundation of our universe and its natural laws contain the information necessary for the creation of life and evolution?
IOW, if life is a guided search program, the space it is searching as it evolves is a space set up by the laws of nature. Thus any information that life discovers and the guiding, active, problem specific information needed to arrive at those targets (according to the [NFL Theorems [link]) already existed within the program of our universe and the search space upon which our universe’s natural laws operate.
“Unless you can make prior assumptions about the ... [problems] you are working on, then no search strategy, no matter how sophisticated, can be expected to perform better than any other”
--Yu-Chi Ho and D.L. Pepyne, "Simple explanantion of the No Free Lunch Theorem", Proc. 40th IEEE Conf. on Decision and Control, Orlando, Florida, 2001.
“This indicate[s] the importance of incorporating problem-specific knowledge into the behavior of the [search] algorithm.”
--David Wolpert and William G. Macready, "No free lunch theorems for optimization", IEEE Trans. Evol. Comp. 1(1) (1997): 67-82.
“The [computing] machine does not create any new information, but it performs a very valuable transformation of known information.”
--Leon Brillouin, Science and Information Theory (Academic Press, New York, 1956).
The understanding that information is never destroyed or created by a computing machine, only transferred is consistent with Conservation of Information theorems, one of which is explained here by Dr. Dembski. The theorem basically shows that, if problem specific information must be added to a search in order to find a target (solve an optimization problem) efficiently – better than chance -- then the information necessary to find that problem specific information is actually just as unlikely or more unlikely and therefore has a higher improbability of being discovered. Thus, the same or larger amount of problem specific information is needed to locate the problem specific information to find the target, ad infinitum.
This is analogous to a library cataloguing system. The catalogue has to be organized in such a non-random way that you can use its organization (problem specific information) to find the information you were looking for at better than random search. Furthermore, that organization (cataloguing system) itself most likely would not have come about through random processes since it would require at least or more than the same amount of problem specific information to find that specific cataloguing organization (information) as that cataloguing system is able to output (measured as an information theoretic probability in bits). COI Theorems show that in order to get better than chance results you must input better than chance results. How do you guarantee better than chance results? According to NFL Theorems, with problem specific information. How do you get problem specific information? Back to COI Theorems: with previous problem specific information ad infinitum, or as observed, from intelligence.
Appeals to non-explanatory “chance of the gaps” in this case are mere hand waving and grasping at straws, and IMO intellectually unfullfiling, especially when the original targets in question are highly improbable, algorithmically complex, specified, not defined by law, yet exert a control aspect on a lawful system and who's operation is highly analogous to observed intelligently designed systems.
Now, we just need to discover how intelligence could have possibly tuned (programmed) and used the universe’s laws of nature to evolve life. In discovering this, we will be discovering how evolution does what it does and the actual amount of fine tuning necessary for different aspects of life. Or is there merely some sort of teleological law inherent in our universe which “just is”? – similar to the intellectually unfullfiling infinite regress of problem specific information. These are some points which can be philosophically debated, while the scientists discover how this intelligence or teleological law designed life.
Quick Intro to the Science of Intelligent Design Theory
First, Id like to include a comment from another blog, Telic Thoughts, to help somewhat clear the fog regarding ID misconceptions.
The comment that I am replying to is centered and colored blue:
MY understanding of the issue (how I view ID) is that ID is interested in discovering how to define the phenomenon we understand as "intelligence" and then to discover the relation between intelligence and other natural systems. Furthermore, ID attempts to discover which phenomenon (if any) is a necessary result of or necessarily a result of the phenomenon known as intelligence. This is where informational systems and information theory comes into play. ID deals with so much more than just biological structures; it deals with complete (informational) systems. So far, the math behind articles such as those of Dr.s Trevor and Abel seems to hold that informational systems will not generate themselves through stochastic processes.
Furthermore, intelligence has the capacity to generate informational systems. To simplify and make a long argument short, add it up and you just might get a correct answer. There is nothing "mystical" or religious going on here … unless you think that intelligence is inherently religious or "mystical."
Well, as far as I am aware, apart from psychoanalysis "why" questions can't be answered scientifically.
As to "when" and "how", those are definitely interesting questions that many scientists have and are working on. Ie: when did the first informational system appear in our universe? As to the how, when it comes to ID there are many opinions … some easier to test than others … all the way from "the universe is programmed to necessarily produce life" all the way to "an outside intelligence had to intervene and seed life." The initial, in my opinion is the more scientific and perhaps testable with mathematical and computer models and simulations, starting with an understanding among many phycisists and cosmologists that the universe itself is the result of the processing of information.
Now, this is where the "who" question comes in. However, I don't know if we can scientifically verify "who" even if a model positing the causal phenomenon as intelligence is the best and most coherent explanation, although that may be further possible ID research (such as the further research which nails the perpetrator after an examination of the forensic evidence posits an intelligent cause). Furthermore, when a person asks, "who," they usually mean that they want a specific name or something. ID Theory can't tell you the name of the guy who designed something, it merely let's you know that there was an intelligent cause to the event you see.
In addition ...
Here is a cycle that is observed within nature: If previously programmed by intelligence, information processors can produce programs which cause laws which, can produce designs and further information processors directly or through evolutionary means. This observation is the description of the creation of evolutionary algorithms to provide solutions to engineering problems.
There are also two other related phenomenon seen in nature: Natural law itself, which is now seen to be a result of the processing of information which causes the program we know as our universe; and the laws produced by the program of life which results from the replication and processing of information. The relation between these two systems -- our universe and life – and the aforementioned cycle is that they follow the cycle from information processor to program to laws to further information processors. Now, we just need to discover the cause. According to the observed cycle mentioned above, programming by an intelligent system is the most probable (and only observed) cause of this type of cycle, instead of accidental generation of information processors (resulting from nothing but chance and necessity) to begin the cycle.
But, the laws can still create other stochastic processes, such as the pattern created by a leaf fluttering to the ground or the position of where a lightning bolt streaks through the sky, without necessarily being designed to create that exact phenomenon. Stochastic processes will result from any set of laws. If there is matter and laws then there will be stochastic processes which are caused by the laws of attraction between matter.
Intelligent Design theorists and proponents operate on the mathematics involved with information theory and on the assumption that the aforementioned observed cycle is a law of nature, thus extrapolating it to related natural phenomenon such as the information processing ability of our universe and life and inferring if a certain type of information exists then previous intelligence is the cause.
As an aside, if inference and extrapolation of observations are not a tool in the scientists’ methodological kit, then a study of past evolution is not science, even though I do think that evolution from primitive replicating information processing biochemical systems has indeed occurred.
With the above understood ...
-ID Theory = Many features of nature are best explained by an intelligent cause because in our experience intelligence is the sole cause of their informational properties. Because ID has been continually verified by observational statistical data ever since human intelligence began to use language and then even more so when human intelligence began designing computer programs, ID is a theory as opposed to being merely a hypothesis. It is a theory which provides the framework for other hypotheses to emerge.
(Upon a basic understanding and definition of the key terms [information, information processor, and intelligence], which I discuss on my blog, a further hypothesis arises within ID theory: Life is necessarily a result of intelligence.)
-Statistical Data for ID Theory: If we take the sample of all information processing systems in which we know the cause, 100% so far have had an intelligent cause.
-more verifying observations (data): The only observations (data) we have of probabilities being overcome in a direction where further information is being created is through designed programs. One way that ID theory, as it relates to biology and evolution, has already been verified is by these evolutionary programs which are intelligently and purposefully designed to converge upon a solution to a specific problem, through the process of controlled random searches acting within the parameters of fine tuned information. Thus, evolutionary algorithms (which in our experience must be designed with a goal with which to generate and converge information) are one major verification of ID theory. Remember this, as I will be discussing it more in depth in latter blog posts.
The comment that I am replying to is centered and colored blue:
I thought that ID ended with the inference that a particular biological structure was designed.
MY understanding of the issue (how I view ID) is that ID is interested in discovering how to define the phenomenon we understand as "intelligence" and then to discover the relation between intelligence and other natural systems. Furthermore, ID attempts to discover which phenomenon (if any) is a necessary result of or necessarily a result of the phenomenon known as intelligence. This is where informational systems and information theory comes into play. ID deals with so much more than just biological structures; it deals with complete (informational) systems. So far, the math behind articles such as those of Dr.s Trevor and Abel seems to hold that informational systems will not generate themselves through stochastic processes.
Furthermore, intelligence has the capacity to generate informational systems. To simplify and make a long argument short, add it up and you just might get a correct answer. There is nothing "mystical" or religious going on here … unless you think that intelligence is inherently religious or "mystical."
To me, the interesting stuff would be when, how, why and of course who designed thestructure, but as I understand it all of that is outside the scope of ID.
Well, as far as I am aware, apart from psychoanalysis "why" questions can't be answered scientifically.
As to "when" and "how", those are definitely interesting questions that many scientists have and are working on. Ie: when did the first informational system appear in our universe? As to the how, when it comes to ID there are many opinions … some easier to test than others … all the way from "the universe is programmed to necessarily produce life" all the way to "an outside intelligence had to intervene and seed life." The initial, in my opinion is the more scientific and perhaps testable with mathematical and computer models and simulations, starting with an understanding among many phycisists and cosmologists that the universe itself is the result of the processing of information.
Now, this is where the "who" question comes in. However, I don't know if we can scientifically verify "who" even if a model positing the causal phenomenon as intelligence is the best and most coherent explanation, although that may be further possible ID research (such as the further research which nails the perpetrator after an examination of the forensic evidence posits an intelligent cause). Furthermore, when a person asks, "who," they usually mean that they want a specific name or something. ID Theory can't tell you the name of the guy who designed something, it merely let's you know that there was an intelligent cause to the event you see.
In addition ...
Here is a cycle that is observed within nature: If previously programmed by intelligence, information processors can produce programs which cause laws which, can produce designs and further information processors directly or through evolutionary means. This observation is the description of the creation of evolutionary algorithms to provide solutions to engineering problems.
There are also two other related phenomenon seen in nature: Natural law itself, which is now seen to be a result of the processing of information which causes the program we know as our universe; and the laws produced by the program of life which results from the replication and processing of information. The relation between these two systems -- our universe and life – and the aforementioned cycle is that they follow the cycle from information processor to program to laws to further information processors. Now, we just need to discover the cause. According to the observed cycle mentioned above, programming by an intelligent system is the most probable (and only observed) cause of this type of cycle, instead of accidental generation of information processors (resulting from nothing but chance and necessity) to begin the cycle.
But, the laws can still create other stochastic processes, such as the pattern created by a leaf fluttering to the ground or the position of where a lightning bolt streaks through the sky, without necessarily being designed to create that exact phenomenon. Stochastic processes will result from any set of laws. If there is matter and laws then there will be stochastic processes which are caused by the laws of attraction between matter.
Intelligent Design theorists and proponents operate on the mathematics involved with information theory and on the assumption that the aforementioned observed cycle is a law of nature, thus extrapolating it to related natural phenomenon such as the information processing ability of our universe and life and inferring if a certain type of information exists then previous intelligence is the cause.
As an aside, if inference and extrapolation of observations are not a tool in the scientists’ methodological kit, then a study of past evolution is not science, even though I do think that evolution from primitive replicating information processing biochemical systems has indeed occurred.
With the above understood ...
-ID Theory = Many features of nature are best explained by an intelligent cause because in our experience intelligence is the sole cause of their informational properties. Because ID has been continually verified by observational statistical data ever since human intelligence began to use language and then even more so when human intelligence began designing computer programs, ID is a theory as opposed to being merely a hypothesis. It is a theory which provides the framework for other hypotheses to emerge.
(Upon a basic understanding and definition of the key terms [information, information processor, and intelligence], which I discuss on my blog, a further hypothesis arises within ID theory: Life is necessarily a result of intelligence.)
-Statistical Data for ID Theory: If we take the sample of all information processing systems in which we know the cause, 100% so far have had an intelligent cause.
-more verifying observations (data): The only observations (data) we have of probabilities being overcome in a direction where further information is being created is through designed programs. One way that ID theory, as it relates to biology and evolution, has already been verified is by these evolutionary programs which are intelligently and purposefully designed to converge upon a solution to a specific problem, through the process of controlled random searches acting within the parameters of fine tuned information. Thus, evolutionary algorithms (which in our experience must be designed with a goal with which to generate and converge information) are one major verification of ID theory. Remember this, as I will be discussing it more in depth in latter blog posts.
Tuesday, November 20, 2007
Zachriel's EA and Dembski's quote
Continuation from here
Hello again Zachriel,
Dr.s Dembski and Marks:
“Such assumptions, however, are useless when searching to find a sequence of say, 7 letters from a 26-letter alphabet to form a word that will pass successfully through a spell checker, or choosing a sequence of commands from 26 available commands in an Avida type program to generate a logic operation such as XNOR [16]. With no metric to determine nearness, the search landscape for such searches is binary -- either success or failure. There are no sloped hills to climb. We note that such search problems on a binary landscape are in a different class from those useful for, say, adaptive filters. But, in the context of the NFLT, knowledge of the class in which an optimization problem lies provides active information about the problem.”
Dembski is saying precisely what he is stating: “Such assumptions are useless when searching to find a sequence of say, 7 letters from a 26-letter alphabet to form a word that will pass successfully through a spell checker ... with no metric to determine nearness, the search landscape for such searches is binary – either success or failure. There are no slopped hills to climb.”
First, what assumptions is he talking about? In the previous paragraph, he refers to the assumptions:
“Within HÄaggstrÄom's familiarity zone, search structures have “links" in the optimization space and smoothness constraints allowing for use of “hill-climbing" optimization. Making such assumptions about underlying search structures is not only common but also vital to the success of optimizing searchers (e.g., adaptive filters and the training of layered perceptron neural networks | [22]).”
So, basically, Dembski and Marks refer to assumptions of links and hill climbing structures in the search space. The point is that merely assuming that there are these link structures within a non-uniform search space , without actual knowledge of these structures incorporated into the search algorithm, does nothing to affect the probability of a *random* search. What is needed is an algorithm that is matched to and programmed to take advantage of a *known* (so that it can provide accurate guidance) non-uniform search structure which actually contains hill climbing optimizations, etc. If the proper active information in the form of the proper type of adaptive search is not applied, there are no slopped hills to climb by *random* search no matter the underlying search structure.
We know that Dembski and Marks are referring to *random* search because in the first quoted paragraph above, he states that “We note that such search problems on a binary landscape are in a different class from those useful for, say, adaptive filters.”
Your algorithm does make use of at least one adaptive filter such as a “stepping stone search” which is defined as “building on more probable search results to achieve a more difficult search” and one instance of stepping stone search is discussed in “Conservation of Information in Search: Measuring the Cost of Success” and a formula for measuring the active information of the discussed stepping stone search function is provided.
Your algorithm is far from the random search of a 7 letter word that Dembski refers to. First, it builds upon the success of previous results (stepping stone search), sorts fit from unfit (also discussed and measurement of active information given in the aforementioned paper) before lesser probable words (any 7 letter words) are achieved, and may actually include other instances of active information. Thus, since it builds off of previous filtered successes, it fits into the category of adaptive searches which Dembski and Marks note are in a separate class than “success or failure” random searches.
Thus, your EA provides an example of Dembski's point ... that *adaptive filters* (which are in a different category than purely random search) must be *properly* applied to take advantage of *known* (as opposed to merely assumed) search space structure -- the non-random dictionary -- in order to provide better than chance performance and arrive at any targeted 7 letter word that will pass the spell checker test. Try the EA on a dictionary that only contains the 7 letter words.
Dembski also states that “Following HÄaggstrÄom [18], let the search space be partitioned into a set of acceptable solutions T and a set of unacceptable solutions T. The search problem is to find a point in T as expediently as possible. With no active information, the distribution of the space is assumed uniform. This assumption dates to the 18th century and Bernoulli's principle of insufficient reason [2] which states, \in the absence of any prior knowledge, we must assume that the events [in ] ... have equal probability." This is equivalent to an assumption of maximum (information theoretic) entropy on the search space [21]. Many criticisms of the NFLT are aimed at the uniformity assumption [28]. Knowledge of a deviation from uniformity, however, is a type of active information that is consistent with the premise of the NFLT.
So, you have separated acceptable solutions from unacceptable solutions by creating a dictionary, but as already mentioned, your search algorithm is designed to take advantage of the non-random aspects of your dictionary. IOW, the algorithm is matched to the solutions. Knowledge of the solutions and the type of algorithm which will work is used to create the correct algorithm which builds upon previous results and also sorts these results in order to achieve lesser probable results in the future at better than chance performance. All of this complementary knowledge is extraneous to the evolutionary search itself and is essential to the success of the search and is thus active, problem specific, guiding information.
You do realize that the most that your EA shows is that the targets for life, and search algorithm to necessarily travel those pathways and to reach the targets are set before the search ever begins. Is that the hallmark of a system built upon a random set of variables?
Now, let’s discuss NFL Theorem a bit:
NFL Theorem states that no method of search (algorithm) will, on average, perform better than any other algorithm [to produce consistently better than chance results] without problem specific information guiding it. The example of finding the Ace of Spades in "Active Information in Evolutionary Search" excellently illustrates the point of the NFL Theorem.
IOW, from what I understand, if you have no knowledge of the search space structure, and you separate acceptable from unacceptable targets, any method (search algorithm) of finding an acceptable target will perform as well as any other method and will on average discover the acceptable target at no better than chance probability. That is, unless active information regarding the search space structure and/or information helping to determine the location of any of the acceptable targets is programmed into the search algorithm in the form of "warmer/colder hints," “stepping stone searches,” “partitioned searches,” “fitness selectors” etc. (Many types of adaptive algorithms are discussed in "Conservation of Information in Search: Measuring the Cost of Success.")
So, what separates an EA from random search? According to NFL Theorem, if the EA is to find a target better on average than random search there must be some problem specific information (knowledge about the targeted problem) guiding it. “Concerning the NFLT, Ho and Pepyne write “unless you can make prior assumptions about the ... [problems] you are working on, then no search strategy, no matter how sophisticated, can be expected to perform better than any other" [15]. According to Wolpert and Macready search can be improved only by “incorporating problem-specific knowledge into the behavior of the [optimization or search] algorithm" [30].
Anticipating the NFLT, Schaefer [24] notes "a learner [without active information] ... that achieves at least mildly better-than-chance performance ... is like a perpetual motion machine." The "prior assumptions" and "problem specific knowledge" required for "better-than-chance performance" in evolutionary search derives from active information that, when properly fitted to the search algorithm, favorably guides the search.” Have you discovered, within your program, the information theoretic equivalent to perpetual motion machines? (Of course, understanding this point, you will see why I am so sceptical re: processes allegedly creating information at averages exceeding probability without guiding, problem specific, active information.)
Now, back to the “knowledge of a deviation from uniformity of the search space” that was briefly mentioned earlier.
The conclusion of “Active Information in Evolutionary Search:”
“Active information, when properly prescribed, successfully guides an evolutionary search to a solution by incorporating knowledge about the target and the underlying structure of the search space. That structure determines how the search deviates from the uniformity assumption of the NFLT. HÄaggstrÄom's "geographical structure[s]," "link structure[s]," search space "clustering," and smooth surfaces conducive to "hill climbing" are examples that reinforce rather that refute the conclusion that intelligent design proponents draw from the NFLT, namely, that the success of evolutionary search depends on the front-loading of active information. Thus, in biology as in computing, there is no free lunch [8].”
Since the NFL Theorem assumes a uniform search space (maximum information entropy), we can conclude that any random search spaces being acted upon by any chosen method of search (search algorithm) will not, on average, arrive at targets at better than random chance, nor will one random set of search space and search algorithm consistently perform better than chance over time. This is consistent with what I understand to be part of the basis for Conservation of Information:
A "learner... that achieves at least mildly than better-than-chance performance, on average, ... is like a perpetual motion machine - conservation of generalization performance precludes it.”
--Cullen Schaffer on the Law of Conservation of Generalization Performance. Cullen Schaffer, "A conservation law for generalization performance," in Proc. Eleventh International Conference on Machine Learning, H. Willian and W. Cohen. San Francisco: Morgan Kaufmann, 1994, pp.295-265.
That can be tested by generating a random set of variables (laws), which causes a random set of functional targets within a search space, and applying any method of search to discover those targets.
According to my understand of the NFL Theorem and Conservation of Generalization Performance, there are two types of results that will not occur:
1. The search algorithm performing consistently better than chance over a lengthy run time.
2. After many shorter runs, of different random searches on random targets, the averages of finding the targets being better than chance.
“Although commonly used evolutionary algorithms such as particle swarm optimization [9] and genetic algorithms [11] perform well on a wide spectrum of problems, there is no discrepancy between the successful experience of practitioners with such versatile algorithms and the NFLT imposed inability of the search algorithms themselves to create information [4, 10]. The additional information often lies in the experience of the programmer who prescribes how the external information is to be folded into the search algorithm. The NFLT takes issue with claims that one search procedure invariably performs better than another or that remarkable results are due to the search procedure alone [1, 3, 13, 14, 17, 19, 23, 25, 26].”
"The NFLT puts to rest the inflated claims for the information-generating power of evolutionary simulations such as Avida [16] and ev [25]. The NFLT gauges the amount of active information that needs to be introduced to render an evolutionary search successful [18]. Like an athlete on steroids, many such programs are doctored, intentionally or not, to succeed [17]."
"Christensen and Oppacher note the NFLT is \very useful, especially in light of some of the sometimes outrageous claims that had been made of specific optimization algorithms" [4].
"Search algorithms do not contribute information to the search, and the NFLT exposes the inappropriateness of such claims."
"The NFLT shows that claims about one algorithm outperforming another can only be made in regard to benchmarks set by particular targets and particular search structures. Performance attributes and empirical performance comparisons cannot be extrapolated beyond such particulars. There is no all-purpose \magic bullet" search algorithm for all problems [5, 27]."
IOW, from what I understand, a non-random, non-uniform search space structure must be coupled to the correct search algorithm, utilizing the proper filter in accordance with prior knowledge of the target, in order to produce better than chance performance when searching for a target.
Now, let’s apply this to real life. Within life, a target is that informational sequence which functions within the living organism. Similarly, if you were to put the dictionary of targets within your EA together using random variables, in order to produce anything worth comparing to the evolution of life, the targets would need to conform to a system of rules and interact with each other to create some type of function. In other words, the targets would need to be algorithmically complex and specified, yet generated by a random set of laws and attained through a random method of search at better than chance. That is what you’d have to demonstrate to even begin to show that life has evolved from non-planned, non- intelligently designed laws and information.
As Dembski and Marks write in “Active Information in Evolutionary Search,” In the field of evolutionary computing, to which the weasel example belongs, targets are given extrinsically by programmers who attempt to solve problems of their choice and preference. But in biology, not only has life come about without our choice or preference, but there are only so many ways that matter can be configured to be alive and, once alive, only so many ways it can be configured to serve biologically significant functions. Most of the ways open to biological evolution, however, are dead ends. It follows that survival and reproduction sets the intrinsic targets for biological evolution.
Evolutionary convergence, the independent evolution of similar features (such as the camera eye of humans and squids), provides a striking example of intrinsic biological targets. Cambridge paleobiologist Simon Conway Morris [20] finds that evolution converges to only a few endpoints. He therefore theorizes that if the evolutionary process were restarted from the beginning, the life forms of today, including humans, would re-evolve. From the perspective of the NFLT, these limited number of endpoints on which evolution converges constitute intrinsic targets, crafted in part by the environment and by laws of physics and chemistry.”
This provides evidence that our universe will necessarily arrive at pre-set living targets which are guided by problem specific, active information at the foundation of its natural laws. This is consistent with any natural teleological hypothesis referencing the origin of our universe and the subsequent evolution of life. Evolutionary algorithms, NFL Theorem, and the Law of Generalization Performance also provide evidence against any assertions that just any random set of information and laws will cause evolution (extremely basic: the, on average, better than chance generation of algorithmically complex and specified information) to occur.
So, your example operates off of some active information and it seems to show the lower capability of evolutionary algorithms. If your EA was to even simplistically model the evolution of life, you need to show that the pathway from useable protein target to the next less probable useable proteins has a high enough probability to be achieved in the amount of time available and this would have to be consistent with known mutation rates (‘x’ mutations per generation, ‘x’ generations per year, 4 x 10^9 years available and largest protein = 26,926 amino acids). Of course, the probabilities between targets would also have to apply to the generation of instructions for assembling machines and systems from those proteins.
Furthermore, your program would also need to model the further evolution of information processing systems, the evolution of the different functions of RNA, the evolution of logic gates, the evolution of complex machinery, the evolution of IC systems (attained through indirect pathways), redundant systems, repair systems, and convergent evolution (the independent evolution of similar features), to name a few systems and results which have evolved within life, not to mention intelligent beings and conscious systems. I wonder how much and the type(s) of fine tuned, problem specific, active information would be necessary in order to evolve all of those results. And "answers" about how our EA’s won’t create such systems just because there isn’t enough time don't impress me. I’m perfectly fed up with anti-scientific, progression crushing, “chance of the gaps” non-explanations. I want knowledge of how something is accomplished. I don’t want “pat” answers about how things “just happen” to self- assemble given enough random variation and time. I want to know *how* in a manner consistent with our knowledge of the flow of information. This truly sounds like some excellent ID research.
The conclusion. Evolution does not create any new information, it only converts it from one program to another -- from the problem specific, active informational structure of our natural laws at the foundation of our universe to the information processing system of life. Enter the Law of Conservation of Information. As well, since evolution generates information at greater than random chance, it must make use of the problem specific information to find the informational targets. Furthermore, evolution by natural selection provides evidence of teleological foresight and programming of the active, problem specific information necessary for the consistently better than chance performance of evolution.
"Search algorithms, including evolutionary searches, do not generate free information. Instead, they consume information, incurring it as a cost. Over 50 years ago, Leon Brillouin, a pioneer in information theory, made this very point: .The [computing] machine does not create any new information, but it performs a very valuable transformation of known information. [3] When Brillouin’s insight is applied to search algorithms that do not employ specific information about the problem being addressed, one finds that no search performs consistently better than any other. Accordingly, there is no magic-bullet search algorithm that successfully resolves all problems [7], [32]."
Next question: How does one information processing system create another information processing system within it (ie: universe creates life)? I predict: not by unguided, random accident. Sounds like some more ID research.
Hello again Zachriel,
Dr.s Dembski and Marks:
“Such assumptions, however, are useless when searching to find a sequence of say, 7 letters from a 26-letter alphabet to form a word that will pass successfully through a spell checker, or choosing a sequence of commands from 26 available commands in an Avida type program to generate a logic operation such as XNOR [16]. With no metric to determine nearness, the search landscape for such searches is binary -- either success or failure. There are no sloped hills to climb. We note that such search problems on a binary landscape are in a different class from those useful for, say, adaptive filters. But, in the context of the NFLT, knowledge of the class in which an optimization problem lies provides active information about the problem.”
Zachriel:
“Dembski has defined the same problem I did above. He claims that without some information built into the algorithm about how words are distributed through sequence space, evolutionary algorithms will be no better than random search. Is this what Dembski is saying? And is he right?”
Dembski is saying precisely what he is stating: “Such assumptions are useless when searching to find a sequence of say, 7 letters from a 26-letter alphabet to form a word that will pass successfully through a spell checker ... with no metric to determine nearness, the search landscape for such searches is binary – either success or failure. There are no slopped hills to climb.”
First, what assumptions is he talking about? In the previous paragraph, he refers to the assumptions:
“Within HÄaggstrÄom's familiarity zone, search structures have “links" in the optimization space and smoothness constraints allowing for use of “hill-climbing" optimization. Making such assumptions about underlying search structures is not only common but also vital to the success of optimizing searchers (e.g., adaptive filters and the training of layered perceptron neural networks | [22]).”
So, basically, Dembski and Marks refer to assumptions of links and hill climbing structures in the search space. The point is that merely assuming that there are these link structures within a non-uniform search space , without actual knowledge of these structures incorporated into the search algorithm, does nothing to affect the probability of a *random* search. What is needed is an algorithm that is matched to and programmed to take advantage of a *known* (so that it can provide accurate guidance) non-uniform search structure which actually contains hill climbing optimizations, etc. If the proper active information in the form of the proper type of adaptive search is not applied, there are no slopped hills to climb by *random* search no matter the underlying search structure.
We know that Dembski and Marks are referring to *random* search because in the first quoted paragraph above, he states that “We note that such search problems on a binary landscape are in a different class from those useful for, say, adaptive filters.”
Your algorithm does make use of at least one adaptive filter such as a “stepping stone search” which is defined as “building on more probable search results to achieve a more difficult search” and one instance of stepping stone search is discussed in “Conservation of Information in Search: Measuring the Cost of Success” and a formula for measuring the active information of the discussed stepping stone search function is provided.
Your algorithm is far from the random search of a 7 letter word that Dembski refers to. First, it builds upon the success of previous results (stepping stone search), sorts fit from unfit (also discussed and measurement of active information given in the aforementioned paper) before lesser probable words (any 7 letter words) are achieved, and may actually include other instances of active information. Thus, since it builds off of previous filtered successes, it fits into the category of adaptive searches which Dembski and Marks note are in a separate class than “success or failure” random searches.
Thus, your EA provides an example of Dembski's point ... that *adaptive filters* (which are in a different category than purely random search) must be *properly* applied to take advantage of *known* (as opposed to merely assumed) search space structure -- the non-random dictionary -- in order to provide better than chance performance and arrive at any targeted 7 letter word that will pass the spell checker test. Try the EA on a dictionary that only contains the 7 letter words.
Dembski also states that “Following HÄaggstrÄom [18], let the search space be partitioned into a set of acceptable solutions T and a set of unacceptable solutions T. The search problem is to find a point in T as expediently as possible. With no active information, the distribution of the space is assumed uniform. This assumption dates to the 18th century and Bernoulli's principle of insufficient reason [2] which states, \in the absence of any prior knowledge, we must assume that the events [in ] ... have equal probability." This is equivalent to an assumption of maximum (information theoretic) entropy on the search space [21]. Many criticisms of the NFLT are aimed at the uniformity assumption [28]. Knowledge of a deviation from uniformity, however, is a type of active information that is consistent with the premise of the NFLT.
So, you have separated acceptable solutions from unacceptable solutions by creating a dictionary, but as already mentioned, your search algorithm is designed to take advantage of the non-random aspects of your dictionary. IOW, the algorithm is matched to the solutions. Knowledge of the solutions and the type of algorithm which will work is used to create the correct algorithm which builds upon previous results and also sorts these results in order to achieve lesser probable results in the future at better than chance performance. All of this complementary knowledge is extraneous to the evolutionary search itself and is essential to the success of the search and is thus active, problem specific, guiding information.
You do realize that the most that your EA shows is that the targets for life, and search algorithm to necessarily travel those pathways and to reach the targets are set before the search ever begins. Is that the hallmark of a system built upon a random set of variables?
Now, let’s discuss NFL Theorem a bit:
NFL Theorem states that no method of search (algorithm) will, on average, perform better than any other algorithm [to produce consistently better than chance results] without problem specific information guiding it. The example of finding the Ace of Spades in "Active Information in Evolutionary Search" excellently illustrates the point of the NFL Theorem.
IOW, from what I understand, if you have no knowledge of the search space structure, and you separate acceptable from unacceptable targets, any method (search algorithm) of finding an acceptable target will perform as well as any other method and will on average discover the acceptable target at no better than chance probability. That is, unless active information regarding the search space structure and/or information helping to determine the location of any of the acceptable targets is programmed into the search algorithm in the form of "warmer/colder hints," “stepping stone searches,” “partitioned searches,” “fitness selectors” etc. (Many types of adaptive algorithms are discussed in "Conservation of Information in Search: Measuring the Cost of Success.")
So, what separates an EA from random search? According to NFL Theorem, if the EA is to find a target better on average than random search there must be some problem specific information (knowledge about the targeted problem) guiding it. “Concerning the NFLT, Ho and Pepyne write “unless you can make prior assumptions about the ... [problems] you are working on, then no search strategy, no matter how sophisticated, can be expected to perform better than any other" [15]. According to Wolpert and Macready search can be improved only by “incorporating problem-specific knowledge into the behavior of the [optimization or search] algorithm" [30].
Anticipating the NFLT, Schaefer [24] notes "a learner [without active information] ... that achieves at least mildly better-than-chance performance ... is like a perpetual motion machine." The "prior assumptions" and "problem specific knowledge" required for "better-than-chance performance" in evolutionary search derives from active information that, when properly fitted to the search algorithm, favorably guides the search.” Have you discovered, within your program, the information theoretic equivalent to perpetual motion machines? (Of course, understanding this point, you will see why I am so sceptical re: processes allegedly creating information at averages exceeding probability without guiding, problem specific, active information.)
Now, back to the “knowledge of a deviation from uniformity of the search space” that was briefly mentioned earlier.
The conclusion of “Active Information in Evolutionary Search:”
“Active information, when properly prescribed, successfully guides an evolutionary search to a solution by incorporating knowledge about the target and the underlying structure of the search space. That structure determines how the search deviates from the uniformity assumption of the NFLT. HÄaggstrÄom's "geographical structure[s]," "link structure[s]," search space "clustering," and smooth surfaces conducive to "hill climbing" are examples that reinforce rather that refute the conclusion that intelligent design proponents draw from the NFLT, namely, that the success of evolutionary search depends on the front-loading of active information. Thus, in biology as in computing, there is no free lunch [8].”
Since the NFL Theorem assumes a uniform search space (maximum information entropy), we can conclude that any random search spaces being acted upon by any chosen method of search (search algorithm) will not, on average, arrive at targets at better than random chance, nor will one random set of search space and search algorithm consistently perform better than chance over time. This is consistent with what I understand to be part of the basis for Conservation of Information:
A "learner... that achieves at least mildly than better-than-chance performance, on average, ... is like a perpetual motion machine - conservation of generalization performance precludes it.”
--Cullen Schaffer on the Law of Conservation of Generalization Performance. Cullen Schaffer, "A conservation law for generalization performance," in Proc. Eleventh International Conference on Machine Learning, H. Willian and W. Cohen. San Francisco: Morgan Kaufmann, 1994, pp.295-265.
That can be tested by generating a random set of variables (laws), which causes a random set of functional targets within a search space, and applying any method of search to discover those targets.
According to my understand of the NFL Theorem and Conservation of Generalization Performance, there are two types of results that will not occur:
1. The search algorithm performing consistently better than chance over a lengthy run time.
2. After many shorter runs, of different random searches on random targets, the averages of finding the targets being better than chance.
“Although commonly used evolutionary algorithms such as particle swarm optimization [9] and genetic algorithms [11] perform well on a wide spectrum of problems, there is no discrepancy between the successful experience of practitioners with such versatile algorithms and the NFLT imposed inability of the search algorithms themselves to create information [4, 10]. The additional information often lies in the experience of the programmer who prescribes how the external information is to be folded into the search algorithm. The NFLT takes issue with claims that one search procedure invariably performs better than another or that remarkable results are due to the search procedure alone [1, 3, 13, 14, 17, 19, 23, 25, 26].”
"The NFLT puts to rest the inflated claims for the information-generating power of evolutionary simulations such as Avida [16] and ev [25]. The NFLT gauges the amount of active information that needs to be introduced to render an evolutionary search successful [18]. Like an athlete on steroids, many such programs are doctored, intentionally or not, to succeed [17]."
"Christensen and Oppacher note the NFLT is \very useful, especially in light of some of the sometimes outrageous claims that had been made of specific optimization algorithms" [4].
"Search algorithms do not contribute information to the search, and the NFLT exposes the inappropriateness of such claims."
"The NFLT shows that claims about one algorithm outperforming another can only be made in regard to benchmarks set by particular targets and particular search structures. Performance attributes and empirical performance comparisons cannot be extrapolated beyond such particulars. There is no all-purpose \magic bullet" search algorithm for all problems [5, 27]."
IOW, from what I understand, a non-random, non-uniform search space structure must be coupled to the correct search algorithm, utilizing the proper filter in accordance with prior knowledge of the target, in order to produce better than chance performance when searching for a target.
Now, let’s apply this to real life. Within life, a target is that informational sequence which functions within the living organism. Similarly, if you were to put the dictionary of targets within your EA together using random variables, in order to produce anything worth comparing to the evolution of life, the targets would need to conform to a system of rules and interact with each other to create some type of function. In other words, the targets would need to be algorithmically complex and specified, yet generated by a random set of laws and attained through a random method of search at better than chance. That is what you’d have to demonstrate to even begin to show that life has evolved from non-planned, non- intelligently designed laws and information.
As Dembski and Marks write in “Active Information in Evolutionary Search,” In the field of evolutionary computing, to which the weasel example belongs, targets are given extrinsically by programmers who attempt to solve problems of their choice and preference. But in biology, not only has life come about without our choice or preference, but there are only so many ways that matter can be configured to be alive and, once alive, only so many ways it can be configured to serve biologically significant functions. Most of the ways open to biological evolution, however, are dead ends. It follows that survival and reproduction sets the intrinsic targets for biological evolution.
Evolutionary convergence, the independent evolution of similar features (such as the camera eye of humans and squids), provides a striking example of intrinsic biological targets. Cambridge paleobiologist Simon Conway Morris [20] finds that evolution converges to only a few endpoints. He therefore theorizes that if the evolutionary process were restarted from the beginning, the life forms of today, including humans, would re-evolve. From the perspective of the NFLT, these limited number of endpoints on which evolution converges constitute intrinsic targets, crafted in part by the environment and by laws of physics and chemistry.”
This provides evidence that our universe will necessarily arrive at pre-set living targets which are guided by problem specific, active information at the foundation of its natural laws. This is consistent with any natural teleological hypothesis referencing the origin of our universe and the subsequent evolution of life. Evolutionary algorithms, NFL Theorem, and the Law of Generalization Performance also provide evidence against any assertions that just any random set of information and laws will cause evolution (extremely basic: the, on average, better than chance generation of algorithmically complex and specified information) to occur.
So, your example operates off of some active information and it seems to show the lower capability of evolutionary algorithms. If your EA was to even simplistically model the evolution of life, you need to show that the pathway from useable protein target to the next less probable useable proteins has a high enough probability to be achieved in the amount of time available and this would have to be consistent with known mutation rates (‘x’ mutations per generation, ‘x’ generations per year, 4 x 10^9 years available and largest protein = 26,926 amino acids). Of course, the probabilities between targets would also have to apply to the generation of instructions for assembling machines and systems from those proteins.
Furthermore, your program would also need to model the further evolution of information processing systems, the evolution of the different functions of RNA, the evolution of logic gates, the evolution of complex machinery, the evolution of IC systems (attained through indirect pathways), redundant systems, repair systems, and convergent evolution (the independent evolution of similar features), to name a few systems and results which have evolved within life, not to mention intelligent beings and conscious systems. I wonder how much and the type(s) of fine tuned, problem specific, active information would be necessary in order to evolve all of those results. And "answers" about how our EA’s won’t create such systems just because there isn’t enough time don't impress me. I’m perfectly fed up with anti-scientific, progression crushing, “chance of the gaps” non-explanations. I want knowledge of how something is accomplished. I don’t want “pat” answers about how things “just happen” to self- assemble given enough random variation and time. I want to know *how* in a manner consistent with our knowledge of the flow of information. This truly sounds like some excellent ID research.
The conclusion. Evolution does not create any new information, it only converts it from one program to another -- from the problem specific, active informational structure of our natural laws at the foundation of our universe to the information processing system of life. Enter the Law of Conservation of Information. As well, since evolution generates information at greater than random chance, it must make use of the problem specific information to find the informational targets. Furthermore, evolution by natural selection provides evidence of teleological foresight and programming of the active, problem specific information necessary for the consistently better than chance performance of evolution.
"Search algorithms, including evolutionary searches, do not generate free information. Instead, they consume information, incurring it as a cost. Over 50 years ago, Leon Brillouin, a pioneer in information theory, made this very point: .The [computing] machine does not create any new information, but it performs a very valuable transformation of known information. [3] When Brillouin’s insight is applied to search algorithms that do not employ specific information about the problem being addressed, one finds that no search performs consistently better than any other. Accordingly, there is no magic-bullet search algorithm that successfully resolves all problems [7], [32]."
Next question: How does one information processing system create another information processing system within it (ie: universe creates life)? I predict: not by unguided, random accident. Sounds like some more ID research.
Subscribe to:
Posts (Atom)
