Efficient Training in Multiagent Reinforcement Learning: A Communication-Free Framework for the Box-Pushing Problem

Abstract

Abstract. Effective coordination is a critical challenge in the design of self-organizing systems (SOSs), particularly in decentralized training where explicit communication incurs high costs. Traditional approaches often rely on inter-agent communication, but this can introduce substantial system overhead. The challenge becomes even more pronounced in multiagent reinforcement learning (MARL)-based systems, where the training process happens in an end-to-end black-box manner. To address this issue, we explore alternative methods to enhance the exploration phase without relying on direct communication, thereby improving search efficiency. Therefore, the shared pool of information (SPI) is proposed in this article, which is a communication-free framework designed to provide agents with structured shared information at initialization. By offering a common foundation for exploration, SPI helps guide group action choices and facilitates more effective decision-making. This approach enables agents to learn coordinated behaviors without bringing the high costs associated with continuous communication. The efficiency of SPI is assessed and validated in the box-pushing problem, a task that requires agents to collaboratively maneuver a box toward a goal while avoiding obstacles. Our findings indicate that SPI accelerates learning, enhances coordination, increases success rates, and optimizes trajectory optimization. These results highlight SPI as a promising and scalable approach for scenarios where communication is costly or infeasible.