Reinforcement Learning for Blackjack

Abstract. This paper explores the development of an Artificial Intelligence system for an already existing framework of card games, called SKCards, and the experimental results obtained from this. The current Artificial intelligence in the SKCards Blackjack is highly flawed. Reinforcement Learning was chosen as the method to be employed. Reinforcement Learning attempts to teach a computer certain actions, given certain states, based on past experience and numerical rewards gained. The agent either assigns values to states, or actions in states. This will initially be developed for Blackjack, with possible extensions to other games. Blackjack is one of the simpler games and the only current game in the SKCards package which needs an Artificial Intelligence agent. All the other games are single player. To test the performance of the Reinforcement Learning agent, several experiments were devised and run. Background Current System The current system is a framework for all card games. There is a reresentation of Cards and gameplay is represented by the interdependant Interfaces Player and UI. The Blackjack Player has an attempted AI implementation, which is a number of nested if-else if-else clauses. This does not work. Blackjack [1] The object is to get upto 21 points in a maximum of 5 cards. This is achieved is by assigning values to each of the cards in the deck: 2-10=Face value; Courts=10; Ace=1/11. The player may then take another card(hit) or end thier turn(stand). Reinforcement Learning Reinforcement Learning learns from rewards for taking a sequence of actions in an evirnoment, based on its knowledge. This will lead to a change and eventually a reward. The agent will accordingly adjust its knowledge. Agents can act based upon values of states or actions in states. It learns based on past rewards for an action, A, and the reward received for action A. For simplicity, we take the average of the previous rewards and update it and also allows the agent to know how good a reward is