Evaluating and Selecting Deep Reinforcement Learning Models for OptimalDynamic Pricing: A Systematic Comparison of PPO, DDPG, and SAC

Abstract

Given the plethora of available solutions, choosing an appropriate Deep Reinforcement Learning (DRL) model for dynamic pricing poses a significant challenge for practitioners. While many DRL solutions claim superior performance, there lacks a standardized framework for their evaluation. Addressing this gap, we introduce a novel framework and a set of metrics to select and assess DRL models systematically. To validate the utility of our framework, we critically compared three representative DRL models, emphasizing their performance in dynamic pricing tasks. Further ensuring the robustness of our assessment, we benchmarked these models against a well-established human agent policy. The DRL model that emerged as the most effective was rigorously tested on an Amazon dataset, demonstrating a notable performance boost of 5.64%. Our findings underscore the value of our proposed metrics and framework in guiding practitioners towards the most suitable DRL solution for dynamic pricing.