Related to the project I'm working on with Philip, I've been studying up quite a bit on population genetics and studies regarding competitive coevolution.

In one paper (Luke and Panait), the authors talk about the best way to evaluate the fitness of a particular individual based on its performance against other individuals.

The two prevalent approaches are:

Duel -- Pair individuals off in a single-elimination tournament, and the dude with the best record is deemed the fittest.

Renaissance -- An individual plays a certain number of games against other individuals. Individuals that have the best performance against a wide variety of opponents are deemed fitter.

Both of these general approaches work reasonably well in Genetic Algorithms, and interestingly enough, the authors point out that most sports employ both in order to determine the "best" team.

Most sports have a regular season, often with two divisions, in which each team plays many other teams, with their performance is being evaluated. This is Renaissance selection. Those with the best record then enter a Duel selection process...the playoffs. Until, of course, there is only one left.

Usually, there's a high correlation between who wins a championship and the who has the best record that year in a given sport. But it probably wouldn't be much fun to just give the big shiny trophy to the team with the best record, and not have any playoffs.

Though that's probably exactly the technique we'll be using to evolve our game players.