[3K1-IS-3] Agents

[3K1-IS-3-03] A Reexamination of Compensation Scheme Design through Distributional Reinforcement Learning

〇Toshihiko Nanba¹, Takahiro Inada¹ (1. Kyoto University of Advanced Science)

キーワード：distributional reinforcement learning, compensation scheme, agency theory

This study compares compensation schemes offered by firms to their salespeople from the perspective of the firm’s profit, using distributional reinforcement learning (DRL). Previous research on compensation design often assumes that employees possess complete information about their environment. In reality, however, salespeople typically do not have full knowledge in advance and instead acquire information through experience, rendering this assumption unrealistic. To address this limitation, we adopt a reinforcement learning approach that not only estimates action values but also their entire probability distribution. Specifically, we employ Quantile Regression Deep Q-Network (QR-DQN) as our primary DRL algorithm and conduct additional simulations using a standard DQN for comparison. Our analysis yields two key findings: (1) a compensation scheme combining a quota and a commission proves optimal under both algorithms, and (2) QR-DQN provides weaker incentives than DQN across all compensation schemes.