A Deep Reinforcement Learning Approach to Dynamic Pricing in Competitive Ride-Hailing Platforms

Abstract

Ride-hailing markets are dynamic, with passenger demand, driver availability, traffic conditions, weather, events, and rival-platform prices changing over time. Static pricing rules and short-term demand-response approaches might not be able to capture these coupled dynamics, and unconstrained price optimization could raise issues of fairness, transparency and affordability. The paper suggests a deep reinforcement learning (DRL) approach for dynamic pricing in a competitive ride-hailing market. The pricing problem is cast as a sequential decision process where the agent observes the market conditions, chooses bounded fare multipliers or fare changes, and is rewarded according to the platform contribution, trip completion, service reliability, driver participation, and customer outcomes. The proposed architecture is composed of state representation via spatiotemporal demand forecast, actor–critic policy with constraints, rival-response model and safety measures to restrict extreme price fluctuation. Training is done in a simulation environment before careful evaluation in the "offline" phase and staged deployment. The research design involves comparing the DRL policy with fixed pricing, a rule-based surge pricing policy, a contextual bandits policy, and a model-predictive control policy for revenue, fulfillment, wait time, cancellation, price volatility, and distributional fairness metrics. The focus is special placed on non-stationarity, multi-agent competition, causal evaluation, and bias in historical data. Rather than purporting to demonstrate the empirical performance benefits, the paper provides a framework for measuring the potential for constrained DRL to enhance marketplace efficiency, while maintaining user trust and regulatory compliance.