Playing DOOM with Deep Reinforcement Learning

Poster
Playing DOOM with Deep Reinforcement Learning

Deep reinforcement learning has been widely used in video game playing. In this project, we trained agents to play scenario ’search and hit’ and scenario ’health gathering’ in game DOOM with deep reinforcement learning model including Deep Q-Learning with experience replay and policy gradient. Additionally we implemented and evaluated the performance of the agents that were trainedwith several different exploration strategies including random policy, epsilon-greedy policy, andBoltzmann policy. To conclude, the agent for the ’search and hit’ scenario achieved almost perfect performance, and the agent for the ’health gathering’ scenario performed far better than the baseline model. We also conclude that randomness in exploration policyenables agents to gather more information about the environment and is importantin training agents.