This reminds me of self-play AI: one of the concepts behind the huge advances in AI performance in games like go and chess. You start with 2 agents that are ini

This reminds me of self-play AI: one of the concepts behind the huge advances in AI performance in games like go and chess. You start with 2 agents that are initially braindead, then record game(s) and then set the agent which lost to train on the game history db, then see if that agent performs better. The cool thing about self-play is that it converges on optimal play w.r.g. to the game history db.