A general framework for reinforcement learning in cognitive architectures

Gustavo Morais, Eduardo Yuji Sakabe, Paula D. P. Costa, Alexandre da Silva Simões, Ricardo R. Gudwin, Esther Colombini

Abstract: Recent advancements in reinforcement learning (RL), particularly deep RL, show the capacity of this paradigm to perform varied and complex tasks. However, a series of exploration, generalization, and adaptation challenges hold RL back from operating in more general contexts. In this paper, we explore integrating techniques originating from cognitive research into existing RL algorithms by defining a general framework to standardize interoperation between arbitrary cognitive modules and arbitrary RL techniques. We show the potential of hybrid approaches through a comparative experiment that integrates an episodic memory encoder with a well-known deep RL algorithm. Furthermore, we show that built-in RL algorithms with different cognitive modules can fit our framework, as well as remotely run algorithms. Hence, we propose a way forward for RL in the form of innovative solutions that integrate research in cognitive systems with recent RL techniques.