I have read about reinforcement learning and more or less understand how it works for a simple game like Tic-Tac-Toe. But I am trying to implement RL and in specific "Sarsa" algorithm in my pacman game, and I don''t really know how to do that, as pacman is quite a different game compared to TTT. I was wondering if anyone could give me a few pointers and guidelines in regard to this problem. Any help would be greatly appreciated.

Well, first you have to have Something to reinforce - that is, some kind of rule for behaviour, that you can punish or reward to make more or less probable in the future.

Samuels'' checker player had a lot of different hardcoded, not very carefully thought out, heuristic rules that were all evaluated and rewarded during game-play, which in the end produced a very decent checkers-player

I guess that in a pacman-game, the rules have to be generated pretty dynamically, since there have to be a lot of different levels - the environment is not the same during all sessions, and rules that are good in one level might not be good in another.

Should the AI learn to play ALL levels (rules are general for ALL levels), or should it learn one level at a time?

What kind of representation of the rules do you want?
Neural network? Classifier system? Decision tree? something else?

-----------------------
Always down - Never out