Learning what to memorize: Using intrinsic motivation to form useful memory in partially observable reinforcement learning

Reinforcement Learning faces an important challenge in partially observable environments with long-term dependencies. In order to learn in an ambiguous environment, an agent has to keep previous perceptions in a memory. Earlier memory-based approaches use a fixed method to determine what to keep in the memory, which limits them to certain problems. In this study, we follow the idea of giving the control of the memory to the agent by allowing it to take memory-changing actions. Thus, the agent becomes more adaptive to the dynamics of an environment. Further, we formalize an intrinsic motivation to support this learning mechanism, which guides the agent to memorize distinctive events and enable it to disambiguate its state in the environment. Our overall approach is tested and analyzed on several partial observable tasks with long-term dependencies. The experiments show a clear improvement in terms of learning performance compared to other memory based methods.