Tune the parameters
We'll do a variation of the genetic algorithm.
We have a population of AI profiles. Then, we draw 4 of them and run a series of random games (no more than 10). There is a loser, a winner and two others. We compute the standard deviation between the genes of the three non-losers, for each parameter. The genes of the loser are replaced by those of the winner, plus a normal error (from the standard deviation, see above).
The genes are:
- the normalized tendency to simulate at random ;
- ... to agree when simulating ;
- the logarithm of the number of iterations ;
- the logarithm of the MCTS parameter.