Toward Gamification Design of Molecular-Level Filters Through Reinforcement Learning of Blackjack

In this paper, we study reinforcement learning of optimal Blackjack strategies on the GPU as an initial step in the gamification design of molecular-level filters for the desalination/depollution problem. We show that through running approximately 30 billion heads-up games, this approach achieves linear speedup greater than 1,000X with computational efficiency exceeding 99%. The data furthermore indicates that the evolved composite Blackjack strategy favorably compares with the known optimal solution.