Explainable Reinforcement Learning Algorithm for Transparent Human-Centric Business Decision Support Systems

Abstract

Many modern businesses are increasingly turning to algorithms to make decisions; however, the lack of transparency within traditional deep reinforcement learning models limits their use in high-risk business contexts. This paper describes an explicit reinforcement learning model, XRL-HBDSS, which is designed to provide an explainable, human-centered approach to business decision support systems. XRL-HBDSS has a proximal policy optimization (PPO) backbone as well as three new modules focused on explaining how the reinforcement learning algorithm makes decisions: 1) a Policy Attention Attribution Network (PAAN) that explains how much importance each feature had at each decision point, 2) a Counterfactual Trajectory Generator (CTG) that generates alternative action sequences, and 3) a Natural Language Explanation Engine (NLEE) that converts the policy gradient into an easily understandable rationale for all stakeholders involved in the decision-making process. XRL-HBDSS is tested on three real-world business tasks: supply chain disruption response, credit portfolio rebalancing, and reducing customer churn. It is compared to six other explainable reinforcement learning algorithms. On average, XRL-HBDSS has a cumulative reward that is 18.7% greater than that of the best of the other competing explainable reinforcement learning models, while reducing the explanation faithfulness error (EFE) to 0.043. In addition, a user study conducted with 124 business analysts indicated that XRL-HBDSS increased decision trust by 34% and reduced perceived cognitive load by 27%. These findings indicate that businesses can achieve complementary transparency and performance when using reinforcement learning for business intelligence.