Decision Making in Multi-agent Systems: from Cooperation to Competition
Abstract:
Supervised machine learning has revolutionized a wide spectrum of decision making systems, where machine learning models, e.g., deep neural networks, are trained using large-scale annotated corpus to solve perception tasks, such as image classification, speech recognition, and machine translation, and then deployed to serve various decision-making tasks, such as medical diagnosis, autonomous driving, and conversational systems. However, this is far from the complete picture and several other challenges remain to be addressed in modern decision making systems: 1) accurate model estimation not necessarily leads to good decision making, as the latter also requires efficiently exploring the action space to combat bandit feedback, i.e., only observe label/feedback of the chosen action, and in the meantime, exploiting the accumulated observations for no-regret decision making over time; 2) what complicates things further is that modern decision making systems typically involve multiple intelligent agents, e.g., AI agents and human, and the interactions among them directly affect the outcomes of decision making and thus cannot be neglected. Depending on the application scenario, we may have altruistic agents that are non-strategic and cooperate to solve a common task and self-interested agents with their own agenda, i.e., they can be cooperative if it benefits them, but can also be strictly competitive when they have antagonistic objectives. My research aims to understand and model such relations between agents in multi-agent decision making systems, as well as developing algorithmic solutions with provable guarantees to address the new challenges that may arise. For problems with altruistic agents, I propose algorithms that enjoy improved regret via cooperative decision making under various challenging environments; for problems with non-cooperative agents, i,e., self-interested agents that are not strictly competitive, I design no-regret algorithms that enable cooperation while satisfying their private interests; and for problems with competitive agents, I design mechanisms to mediate their competition towards preferred outcomes to optimize important societal objectives, e.g., social welfare. The proposed research is expected to build both theoretical foundations and practical impacts over real-world decision making systems.
Committee:
- Aidong Zhang, Commitee Chair, CS/SEAS/UVA
- Hongning Wang, Advisor, CS/SEAS/UVA
- Cong Shen, ECE/SEAS/UVA
- Haifeng Xu, CS/ University of Chicago
- Shangtong Zhang, CS/SEAS/UVA