The article discusses reinforcement learning (RL) in multi-agent systems (MAS), focusing on innate-values-driven behaviors of AI agents. It introduces a hierarchical compound intrinsic value reinforcement learning model, termed IVRL, to describe complex behaviors of multi-agent interaction in cooperation. The model is implemented in the StarCraft Multi-Agent Challenge (SMAC) environment, comparing cooperative performance within three characteristics of innate value agents (Coward, Neutral, and Reckless). The study concludes that rational organization of individual needs leads to better group performance with lower costs.
Publication date: 12 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.05572
