Advancing Deep Learning for Multiagent AI: Mechanisms, Organizations, and Dynamics

Abstract

Effective and safe multi-agent AI holds the promise to solve important problems across domains such as digital markets, multi-robot coordination, human-AI alignment, and modular methods for monolithic learning systems. This thesis aims to advance deep learning methods so that they scale to and provide certifiable alignment guarantees in real-world complexity. Technically, it approaches this goal by encoding structure into deep learning that mediates multi-agent dynamics and enables rigorous analysis. This agenda explores systems of independent agents as well as embodied systems, modeled compositionally as multiple agents. This approach has produced a series of deep learning methods that are the first of their kind, achieving provably aligned and effective multi-agent AI that synergizes insights from computational economics, game theory, reinforcement learning, and robot learning. Specifically, this agenda operates at three macro-to-micro levels of system abstraction, focusing on incentive structures that shape learning outcomes, interaction structures that simplify dynamics, and statistical structures that guide agent learning, respectively. Key contributions are as follows: Incentive structure: At this level, I design principled incentives that make a target behavior an equilibrium to which rational agents converge, avoiding the need to analyze dynamics in detail. To achieve this, I encode a menu-based architecture and function discontinuity into deep learning to address landmark problems in incentive design, including auction and contract theory. Applied to computational economics, this method yields the first provably aligned (truthful reporting) and effective (revenue optimizing) neural mechanism for multi-bidder, multi-item auctions and single-bidder combinatorial auctions. Interaction structure: At this level, I accelerate learning by encoding structures that mediate agent interaction, such as groupings, symmetry, and near decomposability into deep learning. These structures simplify multi-agent dynamics and boost learning efficiency in the control and design of a single embodied agent, and also promote multi-agent cooperation. Statistical structure: At this level, I identify statistics that characterize the dynamics of a system. The problem of agent learning is formulated as modeling and thereby steering these dynamics. I propose a unified approach for dynamics modeling using a generative neural ordinary differential equation that encodes the dependency structure among the statistics. I also study agent modeling using large language models, with the aim of optimizing decision-making policies based on these models. These studies are supported by principled analyses that rigorously assesses the effectiveness of dynamics modeling and policy learning.