Re: JAPH-ing Genetically

[Anonymous Monk]

make a string which, when evaled, gives JAPH! Then you have multiple good answers and a very useful program.

[larsen]

andye is right. Fitness function f must be continuus, so that for two similar individuals x₁ and x₂, f(x₁) is similar to f(x₂).

So, small causes (small random mutations) imply small changes (small variations of fitness).

If you want to apply the ideas of GA to evolve programs, you should explore the so-called field of Genetic Programming; a classic reference are Holland's books

[Anonymous Monk]

You could fuzz hard edges (go - no go) by statistical evaluation (Markov-Chains) with the statistical properties of Perl. Meaning if the program has the same statistical properties as Perl it is slightly fitter than when not, despite it not being valid Perl or if it is valid (which gives a large boost in FITNESS) but doesn't have the statistical properties of Perl it is slightly unfitter.

[Masem]

One of my other examples for Algorithm::Genetic is redoing one of the best nodes, Genetic Programming or breeding Perls, using my module. This node uses bits of legal perl code like "$x += $y;" and "$y=$x;" to make up a long operation; the bits are breed and mutated such that the operation, when evals to a number, and the GA aims to make code that matched a given number. This is a much better GA example, since with a good operations set, there's numerous ways to calculate a given number.

The reason this JAPH one works is that I upped the mutation rate to 10%, such that when it's close but not at the solution, there's a good chance that that one character pertubation is the right one to lock in the right solution.

---

Dr. Michael K. Neylon - [e-mail] || "You've left the lens cap of your mind on again, Pinky" - The Brain

[John M. Dlugosz]

It seems that chemists are using genetic algorithms, too:

link
Designing new enzymes is a difficult task, but what researchers can do is make thousands of trillions of RNA molecules, with the hope that one or a few of them can catalyze the appropriate reactions.

What's especially weird about this case is that the subject of this particular enzime is genetics!

Now if we could make trillions of little perl interpreters and turn them loose on a problem...

—John

[andye]

I'd see the problem with that as being the fitness function. One character could make the difference between complete success and complete failure - so how do you evaluate fitness?

If anyone knows an accepted method for this, I'd find it v.interesting.

Andy.

[Anonymous Monk]

You could evaluate $@ for instance if it contains "syntax error" that means DWIM gives up it is a very low fitness, if it contains "operator expected" it is better and if it contains "not defined" better yet. You get the picture?

[andye]

Problem is, an individual that generates a syntax error might only be one character different to an individual that prints the required string. Or it might be total rubbish.

The first should (for this method to work) be fitter than the second, but I don't know a way to quantify that.

Also, p(syntax error) is very high for constructing a valid Perl string randomly. The chances are that all members of a population would generate syntax errors - at which point they'd all have the same fitness.

A better approach would be to create a baby language safe (or safer) from syntax errors and eval that. Like maybe just arithmetic operators and numbers.

Although - it's just occured to me - you'd be on fairly safe ground with a subset of a highly structured language like Logo. And you could evaluate fitness by checking how similar the figure drawn was to your target figure.

Andy.