Reward Shaping for Deep Reinforcement Learning in VizDoom

Abstract

Reward shaping helps reinforcement learning agents to succeed in challenging tasks when environmental rewards are either sparse or delayed. In this work we propose an approach which combines both information from the game screen and additional information about in-game events to produce an estimation of novelty of the visited states and used behaviors. We use this estimation to motivate the agent to seeking novel experiences and show that our method helps in accelerating learning and reaching better and secures more robust strategies in complex VizDoom scenarios.