Hyungil Ahn, Rosalind W. Picard

Abstract

This paper presents a new computational framework where both ‘the extrinsic reward from the external goal or cost’ and ‘the intrinsic reward from multiple emotion circuits and drives’ play an integral role in learning and decision making. We show that the integration of the intrinsic reward from affect systems can be used for enhancing the efficacy of learning and decision making. In particular, we suggest a model of the affective anticipatory reward that is assumed to arise from the emotional seeking system. Our simulation results for a singlestep choice and sequential multi-step choices show that affective biases from affective anticipatory rewards can be applied for improving the speed of learning, regulating the trade-off between exploration and exploitation in learning more efficiently, and adjusting the weight given to the immediate rewards over the future rewards in obtaining a decision making policy.