Competence-Preserving Retention of Learned Knowledge in Soar's Working and Procedural Memories

Abstract

Effective management of learned knowledge is a challenge when modeling human-level behavior within complex, temporally extended tasks. This paper evaluates one approach to this problem: forgetting knowledge that is not in active use (as determined by base-level activation) and can likely be reconstructed if it becomes relevant. We apply this model for selective retention of learned knowledge to the working and procedural memories of Soar. When evaluated in simulated, robotic exploration and a competitive, multi-player game, these policies improve model reactivity and scaling while maintaining reasoning competence.

Computationally Efficient Forgetting via Base-Level Activation

“…This approach has been empirically evaluated for long-term tasks in the procedural and working memories of Soar (Derbinsky & Laird, 2012). In this paper, we focus on the quality and efficiency of our prediction approach and utilize synthetic data.…”

A cognitive architecture theory of anticipatory thinking

“…Episodic memory stores the experiences of an agent whereas semantic memory stores more general concepts and facts such as geographic information and the meanings of words. These memory systems provide content-addressable memory that is distinct from procedural memory functionality (Derbinsky and Laird 2010). As noted above, the stored knowledge in these memories can only be used in reasoning if it is first retrieved into working memory.…”

40 years of cognitive architectures: core cognitive abilities and practical applications

“…In particular, task decomposition, where the goal is recursively decomposed into subgoals, is a very common form of planning (e.g. Soar [117], Teton [560], PRODIGY [158]). Other types of planning are also used: temporal (Homer [567]), continual (CoSy [210]), hierarchical task network (PRS [124]), generative (REM [381]), search-based (Theo [375]), hill-climbing (MicroPsi [32]), etc Very few systems in our selection rely on classical planning alone, namely OSCAR, used for logical inference, and IMPRINT, which employs task decomposition for modeling human performance.…”

“…The goal is typically not to master the game but rather to use it as a step towards solving similar but more complex problems. For example, Liar's Dice, a multi-player game of chance, is used to assess the feasibility of reinforcement learning in large domains (Soar [117]). Similarly, playing Backgammon was used to model cognitively plausible learning in (ACT-R [460]) and tic-tac-toe to demonstrate ability to learn from instruction (Companions [231]).…”