If this is true (which it might be, I haven’t studied these systems in detail), then an obvious fix is to get the AI to randomly train with ‘weaker’ versions of

If this is true (which it might be, I haven’t studied these systems in detail), then an obvious fix is to get the AI to randomly train with ‘weaker’ versions of itself (perhaps while the strong instance is handicapped) in addition to the latest generation. Several levels of weak opponents should be used, with varying probabilities, to tune the AI’s robustness against real-world, imperfect competitors.