REINFORCEMENT LEARNING-BASED DYNAMIC PRICING IN A STOCHASTIC DEMAND–SUPPLY ENVIRONMENT

Abstract

Dynamic pricing in ride-sharing platforms must balance revenue generation with stable pricing decisions under changing demand and supply. This study aims to develop and evaluate a reinforcement learning-based dynamic pricing policy that maximizes expected revenue while reducing abrupt policy-level price adjustments. A stochastic contextual environment was constructed from 1,000 historical ride records and evaluated using a leakage-safe 70/15/15 train-validation-test split. The agent was trained with Proximal Policy Optimization (PPO) using five discrete price adjustments from -10% to +10%. Expected revenue was combined with a multiplier-based stability penalty, where stability was measured from changes in the price multiplier rather than nominal price variation across heterogeneous rides. Across 30 paired test episodes, the PPO policy achieved a cumulative reward of 103,316.25 +/- 3,243.99 and expected revenue of 103,449.18 +/- 3,241.27, significantly exceeding static pricing (p < 0.001). Relative to rule-based surge pricing, PPO produced statistically indistinguishable cumulative reward (p = 0.808) while reducing multiplier volatility by 24.83%, mean absolute multiplier change by 24.21%, and action switch rate by 10.97% (all p < 0.001). These results indicate that PPO can preserve near-surge revenue while producing smoother dynamic pricing decisions within the simulated environment.