Rainbow-RND: a Value-based Algorithm Augmented with Intrinsic Curiosity

Deep Reinforcement Learning (DRL) is, without a doubt, one of the most promising and exciting research area in Artificial Intelligence (AI). Several approaches have been proposed and improved in a short time to solve different problems. To tackle exploration’s issues, we present in our work a new approach based on curiosity to generate intrinsic rewards. These latter are related to complex environments with sparse rewards, such as the notorious Montezuma’s Revenge and an even more recent and complicated environment named Obstacle Tower. This type of environments requires the agent to generalize its knowledge, learn high-level planning and low-level control. The results of our experimentations showed that a value-based algorithm (such as Rainbow) can be successfully used with a curiosity-based exploration approach (such as random network distillation). This combination effectively performs better in the Obstacle Tower environment.