The below link will allow you to download an article relating to using machine learning to enhance card counting performance. It's largely 'over my head,' but it references Humble, etc. Potential pitfalls include simulations with unrealistically player-favorable penetration as well as the use of "traces" (programming shortcuts in lieu of more involved simulations).

But here is the abstract followed by the link:

Abstract:
This paper investigates a new method to boost Blackjack betting returns. We show betting criteria identified by genetic algorithm significantly outperform standard game theoretic criteria on ten different professional counting systems.

https://www.researchgate.net/publica...tting_Criteria