Reinforcement learning versus data-driven dynamic programming: a comparison for finite horizon dynamic pricing markets

Revenue management (RM) plays a vital role to optimize sales processes in real-life applications under incomplete information. The prediction of consumer demand and the anticipation of price reactions of competitors became key factors in RM to be able to apply classical dynamic programming (DP) methods for expected long-term reward maximization. Modern model-free deep Reinforcement Learning (RL) approaches are able to derive optimized policies without explicit estimations of underlying model dynamics. However, RL algorithms typically require either vast amounts of training data or a suitable synthetic model to be trained on. As existing studies focus on one group of algorithms only, the relation between established DP approaches and new RL techniques is opaque. To address this issue, in this paper, we use a dynamic pricing framework for an airline ticket market to compare state-of-the-art RL algorithms and data-driven versions of classic DP methods regarding (i) performance and (ii) required data to each other. For the DP techniques, we use estimations of market dynamics to be able to compare their performance and data consumption against RL methods. The numerical results of our experiments, which include monopoly as well as duopoly markets, allow to study how the different approaches’ performances relate to each other in exemplary settings. In both setups, we find that with few data (about 10 episodes) fitted DP methods were highly competitive; with medium amounts of data (about 100 episodes) DP methods got outperformed by RL, where PPO provided the best results. Given large amounts of training data (about 1000 episodes), the best RL algorithms, i.e., TD3, DDPG, PPO, and SAC, performed similarly achieving about 90% and more of the optimal solution.