Long-Term Symbolic Memories for Long-Living Learning Agents

We humans, as prime exemplars of long-lived, learning agents, are frequently bombarded with dense and varying torrents of information, including data that is autobiographical (Laird & Derbinsky, 2009; Tulving, 1983), lexical (Miller, 1995), conceptual (Kolodner, 1983), and commonsensical (Lenat, 1995). Despite this deluge of experience, humans do not drown; we push forward, drawing on our knowledge and reasoning abilities to flourish in challenging and novel situations and tasks. The human cognitive architecture efficiently manages large stores of experience and supports precise retrievals, bringing to bear pertinent knowledge to effectively act in dynamic environments (Laird & Wray, 2010). A review of prior psychological and computational work (Derbinsky & Laird, 2010) suggests that this robust behavior is due in part to our multiple, dissociated memory systems, citing significant functional and computational tradeoffs when utilizing a single memory mechanism for different types of learning tasks. While research into cognitive architectures for artificial learning agents typically reflects this dissociation strategy (Langley et al., 2009), significant work must still be done to understand the specific functionalities these memory systems must support to achieve human-level intelligence, as well as how to efficiently implement these mechanisms over long lifetimes. In my thesis work, I seek to improve our functional and computational understanding of two long-term, symbolic memory systems, episodic and semantic, within the context of a general cognitive architecture. Semantic memory stores general facts that the agent knows, independent of the context in which they were originally learned, which can be applied to improve understanding and task performance in numerous, potentially unrelated situations. In contrast, episodic memory stores autobiographical, contextualized agent experience that allows an agent to remember its own past, such as recalling what occurred in similar situations and using that knowledge to decide how to act presently. I will first summarize my work with these memory systems to date, and then continue to my plans for future research.