Deep Reinforcement Learning with VizDoom First-Person Shooter
Dmitry Aleksandrovich Akimov, Ilya Makarov · Munich Personal RePEc Archive (Ludwig Maximilian University of Munich) · 2019
In this work, we study deep reinforcement algorithms forpartially observable Markov decision processes (POMDP) combined withDeep Q-Networks. To our knowledge, we are the first to apply standardMarkov decision process architectures to POMDP scenarios. We proposean extension of DQN with Dueling Networks and several other model-freepolicies to training agent using deep reinforcement learning in VizDoomenvironment, which is replication of Doom first-person shooter. We de-velop several agents for the following scenarios in VizDoom first-personshooter (FPS): Basic, Defend The Center, Health Gathering. We com-pare our agent with Recurrent DQN with Prioritized Experience Replayand Snapshot Ensembling agent and get approximately triple increase inper episode reward. It is important to say that POMDP scenario closethe gap between human and computer player scenarios thus providingmore meaningful justification for Deep RL agent performance.