Modern Foundations of Multi-Agent Learning
Abstract
Rapid advances in AI, in general, and Deep Learning, in particular, have delivered somewhatmuted progress in multi-agent applications, where---barring some notable exceptions---we arestill widely lacking both the theoretical understanding and the optimization andlearningmachinery to address the difficulties that arise in modern (as well as some outstanding classical)challenges.Our proposal squarely aims to address core issues related to multi-agent settings, including 1)identifying tractable solution concepts that are suitable in multi-agent deep learning, andresolving long-standing computational complexity questions; 2) understanding the computationaland statistical complexity of multi-agent reinforcement learning, and how this depends on theunderlying structure of the problem setting and targeted solution concept; and 3) devisingscalable techniques for reasoning and planning under large amounts of asymmetric imperfectinformation in sequential games.Our team comprises world leaders with expertise covering the relevant fields of MachineLearning, Multiagent Systems, Economics, Game Theory, Control, Optimization, and Theory ofComputation. We have an extensive established track record of outstanding theoretical andpractical contributions, including in large-scale multi-agent challenges, such as Diplomacy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 10, 2025
- Source ID
- N000142512296
Entities
People
- Konstantinos Daskalakis
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy