I like this idea, if we train AIs using learning, then fix it, we will (hopefully) get realistic bahaviour but held in a predictable pattern. What I would gladly volenteer to play against AIs in a training environment so they can learn tactics etc. If we had a small arena with lots of stimuli (places to hide, jump on etc.) then this would be totally cool - like the Matrix training programs; sweet. >We wouldn't implement learning in game. I imagine something more like >where we spend a lot of time training our AIs in an arena or something >and then we release them as fixed-knowledge AIs in the game. > >For instance, NNs don't always have to be learning. You can just take >the values in the NN and leave them as-is, and then you have an AI with >all the knowledge of the NN but can't learn. > >Darryl > > > >Andy wrote: > > > >>I only have one question... >> >>what should learning be good for, in our case ??? >> >>