MAD | AI 2020: Evaluating AI Agents to solve the Blackjack problem.

Gabriel Camilleri | University of Malta
Jake Seracino | University of Malta
David Vella | | University of Malta
Jacob Cassar Ellis | | University of Malta
Mark Bugeja | University of Malta

Blackjack is one of the most popular casino games in the world. It involves comparing cards between the players and the dealer. In this research, we implemented a number of AI agents adapted from several machine learning techniques that could solve the Blackjack problem. Each algorithm is designed to approximate the most optimal strategy which dictates what action should be taken given a particular game state so as to maximise winning likelihood. The three algorithms implemented are Q-Learning, an evolutionary algorithm and an evolutionary neural network. Whereas typical studies conducted in the domain focus mainly on three legal actions; hitting, standing and doubling-down, our contribution also considers splitting as this action is allowed in most casino variations of Blackjack. The algorithms mentioned were initially evaluated separately. The Q-Learning algorithm was evaluated on three ordering criterion; the combination which won the most rounds, the combination which lost least, and the combination which had the best amount of net chips. The Genetic algorithm performed five consecutive tests of 100000 rounds, recording the criterion previously mentioned. The Evolutionary Neural Network was tested with different hyperparameters with 5000 epochs each. The aforementioned algorithms are also compared against each other to see which one performs best. Finally, the knowledge learnt by the AI agents was transferred into a Unity-based Blackjack simulation to allow the user to see in real-time the decisions taken by the agent given a particular game state. It is concluded that the GA implemented, approximated a better strategy for blackjack then Q-Learning and ENN.