Game Playing Agent for 2048 using Deep Reinforcement Learning

Reinforcement learning is used in applications where there is no correct approach to solve the problem. Teaching computers to play games is a complex learning problem that mostly depends on the game complexity and representational complexity, which has recently seen increased attention towards this field. This paper presents an appr oach using concepts of reinforcement learning to master the game of 2048. The approach is based on Q learning and SARSA (State-Action-Reward-State-Action) which are the most popular algorithms in the field of reinforcement learning. The design involves the use of neural networks as the function approximation method. Like most deep Q learning models, the heart of any reinforcement learning agent is the reward function. In this paper reward functions are designed to train the model to learn the best playing moves. But in 2048 there are 4 random components, that is, initial configuration of the game, addition of random tiles after very move, exploration of the agent and unavailability of moves. This paper attempts to provide an approach to solve the issue of inherent randomness in 2048. This approach is based on prioritized experience replay where the model trains to learn best game-playing strategy from the experiences collected. With this approach we achieved maximum tile value of 512.