IC4Net: Decentralized Communication for Continual Multi-Agent Learning

Abstract

The purpose of this paper is to address the challenge of continual learning in multi-agent systems by introducing IC4Net, a decentralized communication framework. IC4Net aims to enable agents to coordinate adaptively and efficiently in dynamic environments. Building on IC3Net, it leverages a gated communication mechanism with Proximal Policy Optimization (PPO), allowing each agent to decide when and with whom to share information without centralized control or fixed protocols. The main contributions are: i) formalizing IC4Net, which combines continual learning and decentralized communication; ii) integrating a self-information metric into PPO for gating decisions; and iii) thorough evaluation under concept drift, class imbalance, and noisy channels. Experimental results show that IC4Net maintains robustness and adaptability under concept drift, class imbalance, and noisy communication channels, while reducing unnecessary communication overhead. These results indicate that IC4Net provides an effective solution for decentralized and continual multi-agent systems.