Extracting Episodic Memory Feature Relevance Without Domain Knowledge

Episodic memory provides many important capabilities to a cognitive architecture. One of the challenges of creating a general episodic memory system is to be effective when given no information about the agent’s task. In this paper, we present an effective algorithm for detecting the relevance of the features of episodic memories while only being told when an agent completes a goal. We demonstrate this algorithm using an episodic learning agent in a task that provides the agent with a mix of relevant and irrelevant features. The episodic learner outperforms a variant Q-Learning algorithm that has proven effective in the past.