Optimizing Q-Learning for Dynamic Pricing in Airbnb

Dynamic pricing is the lifeblood of revenue generation in online marketplaces like Airbnb. Reinforcement learning (RL), in particular Q-learning has been extensively investigated in this problem, but the problem becomes challenging since the state space is huge and the learning dynamics are unstable. This work contributes with a novel, multi-agent Q learning, optimization framework for dynamic pricing that additionally combines adaptive open loop exploration-exploitation strategies and economic based reward functions. To show the efficiency of the proposed model we compare with a simple rule-based pricing rule. Experimental results indicate that our Q-learning model generates higher revenues as well as higher dwelling yield with a higher pricing comparatively to heuristics, the difference between revenues of two methods being statistically significant.