Adaptive Motivation System Under Modular Reinforcement Learning for Agent Decision-Making Modeling of Biological Regulation

Abstract

In this paper, an adaptive motivation system under modular reinforcement learning is suggested for agent decision-making modeling of biological regulation. For this purpose, first, main concepts of drives, rewards, action selection under modular reinforcement learning as well as an adaptive priority process are developed. Second, experiments and results are presented and analyzed demonstrating the efficiency of the suggested concepts. Finally, a discussion is given in conclusion with regard to related works. The obtained results demonstrate how the suggested adaptive motivation system can be used by an agent learning (on-line) to select appropriate actions, during a navigation task from a starting position to a goal position (external goal), i.e., in each moving step, in order to reach an external goal as well as to satisfy internal goals (drives such as hunger, thirst, …); predicting a promising result in future to demonstrate how the nature of the interaction (stimulation-drive, social-drive, …) influences the agent behavior.