Difference between revisions of "Grammatical evolution engine"
From Makers Local 256
m (some ideas) |
m (added reference to "Data types à la carte") |
||
| Line 14: | Line 14: | ||
* Generate only well-typed programs from the codons. | * Generate only well-typed programs from the codons. | ||
| + | ** Maybe an extrapolation of some ideas in [http://lambda-the-ultimate.org/node/2700 Data types à la carte] | ||
* Use a stream of codons for each program rather than a bounded array. This would prevent the need to loop back to the beginning during program generation. | * Use a stream of codons for each program rather than a bounded array. This would prevent the need to loop back to the beginning during program generation. | ||
* The idea of reproducing program-generating codons rather than programs is much closer to biology than genetic programming gets. Experiment with making this even closer to biology by trying alternative representations and recombination operations. | * The idea of reproducing program-generating codons rather than programs is much closer to biology than genetic programming gets. Experiment with making this even closer to biology by trying alternative representations and recombination operations. | ||
Revision as of 22:19, 27 July 2009
|
Creator: |
Overview
This is an attempt at grammatical evolution, subsuming the GED project. Grammatical evolution is similar to genetic programming, but it is able to generate code in any arbitrary language. The actual genome representation for a generated program or algorithm is merely a sequence of codons. Each codon represents a choice to make in a BNF for the language the final program is to be written and tested in.
Ideas
These ideas are deltas to the original grammatical evolution research:
- Generate only well-typed programs from the codons.
- Maybe an extrapolation of some ideas in Data types à la carte
- Use a stream of codons for each program rather than a bounded array. This would prevent the need to loop back to the beginning during program generation.
- The idea of reproducing program-generating codons rather than programs is much closer to biology than genetic programming gets. Experiment with making this even closer to biology by trying alternative representations and recombination operations.
- Stream of pairs of codons instead of single codons, with an idea of "dominance" (like dominant vs. recessive genes).
- For example, maybe the larger valued codon is the more dominant.
- The most dominant codon of the pair is the one used in generating the program.
- Try combining a random choice from each pair from each parent instead of crossover operations. Among other things, this might allow interesting recessive traits to survive along with the most fit individuals in the population, giving more variety later in the simulation.
- Stream of pairs of codons instead of single codons, with an idea of "dominance" (like dominant vs. recessive genes).
