One-Shot Supervised Reinforcement Learning for Multi-targeted Tasks: RL-SAS

Abstract

Our ultimate goal is to realize artificial agents, which can be taught and can behave appropriately in volatile environments. Supervised reinforcement learning (SRL) will play a crucial role in this endeavor as SRL enables agents to function in situations that partly deviate from what has been taught. Currently reinforcement learning (RL) is typically implemented for single tasks, which restricts teaching plural behavioral sequences. Herein we introduce a SRL scheme, which exploits explicit state-action lists to facilitate reuse of learned behavioral sequences. By combining the constructed learning system with a standard RL algorithm, the system could solve a problem in one-shot for the supervised portions and use RL to compensate for the unsupervised portions.