The purpose of this project is to use reinforcement learning to train an agent to play blackjack. Two different representations of the state space will be considered, one which is based on representing the state as the set of cards and the dealer’s card sum, this will be referred to as the ”hand” state-space and one that represents the state as the agent’s and dealer’s card-sum. The card-sum representation is implemented in some standard approaches to Blackjack, but is somewhat flawed (as will be discussed later). Both ”On policy Monte-Carlo learning” and ”Q-learning” will be considered.