주인-대리인 문제를 위한 강화학습

Abstract

Principal-agent relationship is established when one delegates decisions that affect one's interests to another, and it exists in various forms in our society. Therefore, studying the principal-agent problem, in which the principal for agent makes decisions against the principal's interests due to conflicts of interest or asymmetry of information, is very crucial in the process of analyzing phenomena or making decisions that occur in our society. The main topic in the principal-agent problem is what incentive structure the owner should present to the agent. The optimal incentive structure problem in the standard model is a bilevel optimization problem, and studies have been conducted on solving problems with restrictive assumptions due to its complexity. In this study, a reinforcement learning approach is proposed to derive the optimal incentive structure under more relaxed conditions than previous studies. This approach derived not only better solution in the same situation than existing algorithms, but also reasonable solutions to the problem where existing algorithms cannot be applied.