As perceived by an individual user, our virtual Tron plays differently every time, as agents are switched before each game. Simultaneously, due to the evolutionary process, the overall level of play increases. This heterogeneous behavior is part of the success; the strength of an individual strategy is valid only given the collective behavior of all other agents -- if the same strategy were used over and over, humans would quickly adapt and the artificial opponent would appear boring and repetitive. This amounts to a mixture of experts architecture , but here the mixture is a consequence of evolution, of agents exploiting different niches created by their opponents. No single strategy can dominate, as long as there are humans who learn to respond to it, bringing down its fitness, making the way for new ones to take its place.
An interesting question for further study is to look for those emergent niches: is it possible that some agents may have adapted to specific subsets of humans? It is conceivable for example, that some agents are better against people with certain styles -- novices for example, or aggressive players.