Think globally - act locally : scalable networked artificial intelligence
Abstract/Contents
- Abstract
- Machine learning aims to relegate many decision-making processes to agents (machines). In applications like autonomous vehicles, communication networks, and robotics, the agents interact in a networked environment, where the decisions of one agent affect the other agents and their learning. As game theory predicts, this interaction can lead to inefficient outcomes for all agents involved. However, machines follow programmatic objectives and protocols that, unlike humans, are not limited by selfish interests but by information and resources. This modern paradigm calls for new tools to design efficient multiagent protocols for distributed and cooperative agents, which is the subject of this thesis. First, we present a distributed learning algorithm that can maximize the sum of rewards in a general game (for a concave objective), based on reward values alone (i.e., "bandit feedback") and limited communication. We then show that in special cases, we can design algorithms that use less (or no) coordination by leveraging the structure of the application. This is done by targeting the Nash Equilibrium (NE) of the game instead of the optimal actions directly. For distributed energy allocation, we show that although poor NE may exist, best-response dynamics (BRD) converge with high probability to an asymptotically optimal NE (in the size of the network). The idea is to look at a random game (i.e., network) rather than the worst-case one. When poor NE are unavoidable, we propose a manager that can learn to steer the inefficient NE toward an efficient point. This game control offers another tool to design efficient and mostly distributed multiagent systems since the manager does not need to know the game and only observes high-level feedback. Finally, we take a step towards dynamic multiagent optimization by considering the accumulated players' rewards over time instead of a static objective. With such a dynamic objective, players can learn to share resources over time, which can significantly increase their average reward.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2023; ©2023 |
Publication date | 2023; 2023 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Bistritz, Ilai |
---|---|
Degree supervisor | Bambos, Nicholas |
Thesis advisor | Bambos, Nicholas |
Thesis advisor | Ashlagi, Itai |
Thesis advisor | Van Roy, Benjamin |
Degree committee member | Ashlagi, Itai |
Degree committee member | Van Roy, Benjamin |
Associated with | Stanford University, School of Engineering |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Ilai Bistritz. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2023. |
Location | https://purl.stanford.edu/xy812tq0561 |
Access conditions
- Copyright
- © 2023 by Ilai Bistritz
Also listed in
Loading usage metrics...