TESTING BEHAVIORAL THEORIES OF MOTIVATION IN ATARI AGENTS

Maria Gabriela Valeriano, Esther Luna Colombini

Keywords: Motivation, cognition, reinforcement learning, behavioral evaluation

TL;DR: We introduce drive and salience motivational modulations in model-free RL and propose behavioral metrics to compare their cognitive effects.

Abstract:

Motivation plays a central role in biological cognition by shaping attention, valuation, and action selection under internal state constraints. However, in most reinforcement learning systems, motivation is reduced to a fixed external reward signal.
In this work, we investigate how different motivational theories from behavioral neuroscience alter agent behavior when embedded into the decision process of a learning agent. We implement motivational modulation mechanisms on top of a model-free reinforcement learner and evaluate their behavioral consequences in interactive environments. Our goal is not to improve benchmark performance, but to study how distinct motivational formulations lead to different patterns of cognitive behavior, exploration, and cue responsiveness. This provides a computational bridge between motivational theory and cognitive decision-making in artificial agents. This work is currently in progress and full experimental results are not yet available. Nevertheless, we argue that the proposed computational formulation, experimental protocol, and theoretical integration between motivational neuroscience and reinforcement learning provide a useful basis for discussion and early feedback within the cognition and motivation research community.

Paper Type: New Short Paper

Submission Number: 69