While that I felt that my C++ NEAT implementation technically works in most areas, I still had to fix the problem handling complexifying. Complexifying is the process how keeping innovations when developing the neural networks. So far, I have only tried to blindly implement the solution described by this online paper https://www.cs.cmu.edu/afs/cs/project/jair/pub/volume21/stanley04a-html/node3.html#SECTION00033000000000000000.
The papers suggestion is making my code way slower though, and also, the paper is from 1999. This makes me wonder if there exists better genetic algorithms today. I will probably look into other ways to save innovation. But if the extra slowness is disregarded I guess then main remaining problems include deciding on specific parameters used in the code as well as adding more ways to create new generations. Examples being different elitism methods or roulette-selection. Finally when all these sub-issues are resolved good enough I must get better acquainted with behavor trees, which will play a large role in my thesis work.
No comments:
Post a Comment