GA Chess

There are computer models that attempt to simulate biological evolution, and they are so vastly oversimplified and divorced from the biological reality they attempt to imitate that claims made on their behalf should be considered very carefully. This brings me to the subject of Genetic Algorithms, often used as definitive proof by Darwinists in debates.

Both the computer simulation AVIDA and biological evolution are instances of Darwinian processes. The only ingredients required for a Darwinian process are replication, heritable random variation, and fitness selection. In both, organisms are able to reproduce and pass their characteristics to their offspring. In both, random mutations arise which affect the organisms’ ability to survive and produce offspring. In both, selection pressures favor some varieties and penalize others.

The key difference is that the Darwinian processes in biological evolution are supposed to be completely blind/dumb/purposeless while--as will be shown below--AVIDA is purposely designed to evolve complexity in a selected environment. Bill Dembski would term this the incorporation of "active information". These are the intelligently designed aspects of the GA that allow it to function and find the target within the search space. But while the search is guided by the simulation parameters the search pathways are not predefined and the end goals are more generalized and not explicit. So AVIDA is only blind/dumb to a certain extent. There are also other helpful functions in AVIDA that do not occur in nature.

I think of fitness functions as a “funnel” that must be properly constrained in order to provide results. The design of this funnel must be balanced; it can either be too constrained or not constrained enough. The programmer’s goal is to find a balance by which the stated goal can be reached. Based upon the results of researchers, there really isn’t such a thing as a “generic” GA program which can solve anything thrown at it–each program has to be designed to fit a purpose.

You must be certain your understanding of the definition of "irreducible complexity" is not faulty. While it's true that AVIDA does not mimic real-life biology (and the authors do not deny that), it does show that an IC system can evolve in tightly constrained environments under certain conditions of replication, variation, and selection. This is important, as some ID supporters seem to regard "irreducibly complex" as tantamount to "unevolvable in principle". This is not a problem since IC primarily deals with DIRECT Darwinian Pathways and always has. Behe has always stated that INDIRECT Darwinian pathways are another matter. And we're talking about Darwinian processes in Biological Reality, which is not nearly as constrained towards an end goal…

We must also consider the level of complexity: the number of components involved. Behe has already noted that IC structures with a small number of components (2-4) may form. But what about IC structures that are composed of tens or hundreds of interlocking components? That is the real problem ignored by Darwinists.

The key contention with Intelligent Design is whether Random Mutations + Natural Selection can produce Complex Specified Information. (Note that some people would prefer "random variations" or "non-foresighted mechanisms" instead of "random mutations", but for the purposes of this article I will use "random mutations".) I do not think any serious ID proponent doubts that Intelligence + Random Mutations + Natural Selection can indeed produce results. The key point is that simulations like AVIDA are set up precisely so they can produce results...NOT necessarily that they strictly follow nature as a guideline. The question is what can RM+NS do under the much broader constraints of nature without intelligence (active information) being involved.

In the “The Evolutionary Origin of Complex Features,” published in Nature in 2003 by Lenski, the selective forces that have 100% probability affixed are those for various simple binary arithmetic functions, which are ultimately used to build the “equals” (EQU) function, and for the EQU function itself. What’s more, the more complex the function, the greater the reward given to the digital organisms for it. There is no analogy for such selective forces in nature. Nature doesn’t care whether something is more or less functionally complex; it only cares whether it can survive in a particular environment. And what happens when no step-by-step rewards are given for functional complexity? An article on AVIDA in Discover magazine last year (Feb. 2005) stated, “When the researchers took away rewards for simpler operations, the organisms never evolved an equals program.” By building rewards into the system — i.e. providing a highly constrained fitness function — the programmers gave the system a purpose. Hence its creative power:

dynamics.org/Altenberg/FILES/LeeEEGP.pdf

“Both the regression and the search bias terms require the transmission function to have ‘knowledge’ about the fitness function. Under random search, the expected value of both these terms would be zero. Some knowledge of the fitness function must be incorporated in the transmission function for the expected value of these terms to be positive. It is this knowledge — whether incorporated explicitly or implicitly — that is the source of power in genetic algorithms.”

Let’s say I have a Chess GA program. Assume abiogenesis and start off with an AI script that recognizes the environment (the chess board) and knows how to move the pieces (survive in the environment) and has a certain basic strategy. At startup this script is duplicated many times without any mutations. The scripting system making up simulated life cannot be abnormally simplistic, like with AVIDA, and the scripts must have the ability to replicate themselves. The functionality for replication must not be protected. The replication process is capable of producing AI scripts that no longer recognize how to play certain elements of chess or they cannot compile at all (death). As in, replication is not limited to producing fully functional chess strategies. Unfortunately the rules of chess are static so the environment doesn’t change.

Now let’s say I applied a very broad constraint in my fitness function: if the script still retains the ability to compile (aka play chess) then it survives. “Old” scripts eventually die. “Lower lifeforms” are afforded a niche where they thrive instead of arbitrarily being eliminated in favor of “higher lifeforms” based upon a constrained process. As in, winners of games get duplicated more often, and with a larger population comes more processor time for this subsection of the population, but losers are not necessarily eliminated in an arbitrary fashion. They just need to be capable of basic survival. Thus a group of “winners” may eventually be modified to the point they start losing horribly or they split off.

That’s it.

Now let’s say I applied a very narrow constraint in my fitness function: the script must not only compile but it must win its game in a small number of moves in order to survive. This is tantamount to the environment being overly hostile.

I wonder what you could expect from these approaches? Personally I would expect that the population of Model 1 might manage to maintain stasis for a while before perishing. I fully expect the population to die off completely in Model 2.

Now if I desired the optimal Chess GA program I would first make it so that during replication that the randomizer function would NEVER produce a script that couldn't compile. The random changes made would always be valid modifications to a chess strategy. More importantly, the fitness function would be highly constrained to only consider strategies that win the most games. Elimination (death) would not depend on general survivability in the environment of chess but instead would be constrained by my goal. While, yes, the pathways through the search space aren't predetermined the overall intelligent goal is to search out the end result of the best chess strategy by constraining the search to a pathway where only the strategies that win the most are even considered. This, of course, is not like biological reality but it will produce my desired goal of finding the best Chess strategy.

Now all three models are Darwinian processes using replication, heritable variation, and fitness selection. But only the third model is likely to produce any usable results since I've designed the environment and constraints with a goal.

The implications of broad constraints in nature to the debate over ID should be clear. If a constraint is tightly controlled a desired result can be gained since only a certain portion of the search space, a target, is tried. Of course, in general the constraints set by nature are nothing like an optimal Chess GA program--nature's constraints are typically very broad and the environment can change dramatically. At the same time an environment in nature can be so constrained that everything within it is eliminated. In nature there must be a balance for survival. An intelligently designed program would never set the constraints so narrowly that everything was eliminated. Nor would it set the constraints so broadly since the desired result would never be reached (this being close to a random search, the probability is not 0 but it's at least highly implausible).

I'll give an example of why broad constraints are such an issue. Richard Dawkins explains the evolution of the bat wing in The Blind Watchmaker: A squirrel-like creature got a mutation that put a flap of skin in its armpit which aerodynamically helped break its fall when it tumbled out of a tree. Its friends without the mutated armpit flaps broke their necks and died when they fell out of trees. Bigger mutated flaps helped decelerate the creatures from higher altitudes, and so on until we have proper wings.

Yes, he was dead serious. But the point is that he recognizes the further need for a narrower constraint since the general constraints defined by the environment wouldn't help the development of wings. There needs to be a refinement of the fitness function, in this case elimination of squirrels which don't have flaps of skin in their armpits. Of course, this story doesn't take into account how often squirrels do indeed fall to their deaths...so this narrower constraint might be a very weak factor indeed.

Some might object that the GA Chess example is too simplistic to produce results that can happily be called conclusive. Despite the high level of abstraction and tightly refined functions the authors of AVIDA defend it as a valid example of Darwinian principles in general. So the limitations we discover should in principle extend to nature. And if it is the case that nature isn't properly balanced in order to funnel the search as it appears to be then we have a very interesting bit of evidence. But even ID proponents won't be completely happy until a more realistic simulation is run.

Update: Went through and zapped grammar/spelling issues.

Update2: Added several more paragraphs for a bigger punch at the end.