Publication: A Distributed Framework for Learning Agent Rationality
Files
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Access Restrictions
Abstract
In this thesis, we present a method for learning convex optimization model inputs from observed data in multi-agent systems using a novel distributed approach that is generalizable to an extensive range of agent, constraint, and objective function settings. Using recently proposed methods for differentiating the solution map of a convex optimization problem with respect to its parameters, we develop a centralized approach and a distributed, agent-based framework for recovering agent parameters—which can be understood as indicators of agent "rationality"—from behavioral observations. We test these models on a dynamic network energy management example and an instance of an electric vehicle charging control game, and we achieve similar performance between the centralized and distributed learning methods.