Unraveling the Complexity of Memory in RL Agents: an Approach for Classification and Evaluation

Researchers propose a standardized framework for classifying and evaluating memory capabilities in reinforcement learning agents, drawing from cognitive science concepts. The paper addresses confusion around memory terminology in RL and provides practical definitions for different memory types along with robust experimental methodologies.

• →Current RL research lacks unified methodology for evaluating agent memory capabilities, leading to inconsistent comparisons.
• →The paper provides precise definitions of memory types including long-term vs short-term and declarative vs procedural memory.
• →A standardized experimental framework is proposed to objectively assess different classes of agent memory.
• →Memory incorporation is essential for RL tasks requiring past information use and adaptation to novel environments.
• →Empirical experiments demonstrate the importance of following proper methodology when evaluating agent memory capabilities.