Hello everyone,
Just to give you a quick update: I finished a prototype for running NEAT
with SuperTux.
The prototype is capable of learning to solve about 70-80% of "Welcome
to Antarctica" pretty quickly, in about an hour maybe.
The code is still hosted at
https://github.com/ChristophKuhfuss/supertuxNEAT/ .
If you want to try it out yourself: compiling should work as usual, I
modified "--help" to show some evolution-specific command line arguments.
I suggest using a sensor grid size of 10-15 with a padding starting from
20. Bigger grids seem to reduce learning capabilities of the neural
networks (overfitting maybe?).
Please note that this is still an early prototype.
I would love to hear your opinion on that if you are interested in the
project and have the time to check it out.
Best regards,
Christoph Kuhfuß
_______________________________________________
Supertux-Devel mailing list
Supertux-Devel@lists.lethargik.org
http://lists.lethargik.org/listinfo.cgi/supertux-devel-lethargik.org
