Neural-Fitted Temporal Difference Learning to Learn to Play Connect Four

In this thesis, neural-fitted temporal difference learning, a form of reinforcement learning, is used to learn to play the game of Connect Four. Seven different artificial players are compared, using five different neural networks. While the first network only uses the basic board state as input, the larger ones also use specific features: rows of two, three and four in the different rows of the board state. It is shown that these features dramatically improve the performance of the agent. Furthermore, two different exploration strategies are used: Boltzmann with constant temperature and epsilon-greedy. The results show that epsilon-greedy gives the most stable result. Finally, the smallest network was given the same number of hidden nodes as the largest network, showing that adding hidden nodes does not improve the score of the system.