Doğay Kamar, Nazım Kemal Üre, and Gözde Ünal addressed the problem of efficient exploration in reinforcement learning by proposing GAN-based Intrinsic Reward Module(GIRM). The aim of GIRM is to learn the distribution of the observed states and to compute an intrinsic reward to lead the agent to unexplored states. They evaluated GIRM in Super Mario Bros and Montezuma's Revenge and show that GIRM is capable of exploring efficiently.
Conference: ICAART 2022-International Conference on Agents and Artificial Intelligence