A second experiment attempts learning by coevolution and hill-climbing. An initially blank player is slightly mutated and plays a tournament with these `descendants'. The winner becomes the father of a new generation.
An evolutionary path is thus started. But this method is incapable here of the success it achieves on other domains. The reason is that being better than a close relative gives no information at all on the performance against radically different strategies.
Whereas having an absolute fitness function would allow this incremental methods to climb towards a good solution, Tron has a complex landscape where successful strategies can differ a lot from each other.
This family evolution can climb one of these peaks, but will never get to compare with distant configurations.
Live and let live: An individual is
evolving through small mutations. The
hill-climbing algorithm does not find
general strategies. Instead, the
player in this picture will have a
long life, provided that the other is
doing the same