Pretty sure this is an issue with either what rewards I am giving them or my config, but I have been stuck on this for over a week now. Making a pong AI where, currently, the agents only rack up points based on hitting the ball. I tried giving them points for scoring points or winning the game. I also tried taking away points for losing the game or having points scored against them. Neither of these seemed to work either, and I figured since it's two AI's playing against each other the difference of having them get rewards for scoring points etc wasn't really needed since the main thing is for them to hit the ball. If they hit the ball 100% of the time, it won't matter, they'll win naturally by doing that. As for the potential config issue, I've tried to research what would be good values for it, I've looked at the config files for the example projects, but I can't get much to change with it. Any assistance would be appreciated. Kind of crazy the agents can go from 200 points to 0 in no time at all lol
#Pong AI learning then becoming stupid again
wispy epoch
finite warren
what/who is the agent playing against?
if it's the same agent playing both sides then it's going to get confused, you will want to do self play or use some hardcoded AI to train against
wispy epoch
I will try that. I had thought of that, but I didn't think it would make that much of a difference
wispy epoch
I turned one of them to Heuristic mode and just let it sit there, training the other one against a non moving target. Attempting to use that model on one side and train the other agent now that they both aren't just flailing about... but it ends up doing the same thing, earning a lot of points, then dropping off. Is there a point where I can ever pin the two agents against each other to learn? I figured as long as one side isn't trying to learn it would be ok, but that isn't working still