Intrinsically Motivated Decision Making for Situated, Goal-Driven Agents

Abstract

Goal-driven agents are generally expected to be capable of pursuing simultaneously a variety of goals. As these goals may compete in certain circumstances, the agent must be able to constantly trade them off and shift their priorities in a rational way. One aspect of rationality is to evaluate its needs and make decisions accordingly. We endow the agent with a set of needs, or drives, that change over time as a function of external stimuli and internal consumption, and the decision making process hast to generate actions that maintain balance between these needs. The proposed framework pursues an approach in which decision making is considered as a multiobjective problem and approximately solved using a hierarchical reinforcement learning architecture. At a higher-level, a Q-learning learns to select the best learning strategy that improves the well-being of the agent. At a lower-level, an actor-critic design executes the selected strategy while interacting with a continuous, partially observable environment. We provide simulation results to demonstrate the efficiency of the approach.