We review an experiment in co-evolutionary learning of game playing where we show experimental evidence that the straightforward composition of individually learned models more often than not results in diluting what was earlier learned and that self-playing can result in reaching plateaus of uninteresting playing behavior. These observations suggest that learning cannot be easily distributed when one hopes to harness multiple experts to develop a quality computer player and reinforce the need to develop tools that facilitate the mix of expert-based tuition and computer self-learning.
展开▼
