UCLA

Applying deep reinforcement learning to multi-agent environments has become a popular trend. However, creating adaptive agents that perform well in dynamic, complex settings remains challenging. Key difficulties include (1) scaling with the number of agents, as complexity grows exponentially with each additional agent, (2) adapting to new environments, where agents need to leverage past experiences, (3) cooperating with unfamiliar agents, since agents trained in fixed groups must interact effectively with unseen peers, and (4) operating within multi-modal input scenarios, beyond simple vector inputs.We identify the limitations of current approaches in addressing these challenges and propose novel methods to overcome them. The contributions of this thesis include the following:

Evolutionary Population Curriculum (EPC): A training approach that enables agents to gradually adapt from small to large groups through a mix-and-match strategy, enhancing scalability.Social Gradient Fields (SocialGFs): A novel gradient-based state representation for multi-agent reinforcement learning, leveraging denoising score matching to learn social dynamics from offline samples. This adaptive representation allows agents to exhibit diverse behaviors. Inverse Attention Network: A mechanism that models agents’ Theory of Mind (ToM) by inferring attentional states based on observations and prior actions, refining attention weights to improve decision-making. Multi-Modal Multi-Agent Benchmark in Minecraft (TeamCraft): A comprehensive benchmark designed to highlight the limitations of current Vision-Language Models (VLMs) in handling complex, dynamic multi-agent environments, setting a new standard for future research in multi-agent systems.

These contributions advance the field by providing scalable and adaptive solutions to fundamental challenges in multi-agent systems.