Combining Coaching and Learning to Create Cooperative Character Behavior

We present a concept for developing cooperative characters (agents) for computer games that combines coaching by a human with evolutionary learning. The basic idea is to use prototypical situation-action pairs and the nearest-neighbor rule as agent architecture and to let the human coach provide key situations and his/her wishes for an associated action for the different characters. This skeleton strategy for characters (and teams) is then fleshed out by the evolutionary learner to produce the desired behavior. Our experimental evaluation with variants of Pursuit Games shows that already a rather small skeleton –that alone is not a complete strategy– can help solve examples that learning alone has big problems with.