Application of Multi-Agent Reinforcement Learning to RLSS Material Circulation Control System

A Regenerative Life Support System (RLSS) is a system that establishes self-sustained material recycling and circulation within a space base on the Moon or Mars. This is a large-scale and complicated system comprising a lot of components such as humans, plants and material circulation system. A RLSS contains many factors with uncertainty, such as dynamics of plants and humans, and failure and performance deterioration of devices. In addition, a RLSS is a large-scale and complicated system extending gradually. An environment with uncertainty or a large-scale and complicated system may not be properly addressed by a centralized system. In particular, such a system cannot always gather accurate information in one center in a frequently shifting environment, thus appropriate processing may be difficult. Therefore, we tried autonomous decentralization of information or decision-making using a Multi-Agent System (MAS). This report discuss the designing of a RLSS material circulation control system using a MAS and a method in which a MAS acquires cooperative action in a bottom-up approach by means of computer simulation. So far, we have confirmed the effectiveness of this method for a RLSS material circulation control system and proved that automatic acquisition of cooperation rules with the autonomous learning among agents are enabled.