Co-evolution

The main reason for introducing co-evolution was to provide gradual selection pressure. Instead of just evolving one population against a fixed fitness measure I instead setup two populations, A and B, each initialised with random individuals. I then picked a random individual from A and evolved B until it produced a player that could beat the individual from A. This individual, now the strongest from B, was used to evolve A. Once some player from A could beat it, I swapped again and so forth:

1. Select random individual from A, Best of A, to get things started.
2. Evolve B until it can beat Best of A and thus select Best of B.
3. Evolve A until it can beat Best of B and thus select Best of A.
4. Repeat from step 2.

This meant that small steps were continually taken to improve the AI and, in theory, I should have ended up with much stronger players than when evolving against Mr Random, as per my previous post on rule inference. That was not the case however - when playing against the best individuals it was evident that they were pretty poor, not much stronger than the ones evolved against Mr Random.

I tried much larger populations and leaving it running for several days on more powerful PCs but to no avail, so I decided to analyse the evolution process and the individual generations in more detail...