报告题目：A reinforcement learning approach for hotel revenue management
报 告 人：陈吉
We develop a new data-driven approach for hotel revenue management. In this approach, the dynamic capacity allocation problem for multiple class customers is solved in two steps: First, a recommended average price is computed, and then the capacity allocation decision is made based on the average price. The recommended average price is computed with a reinforcement learning algorithm. The capacity allocation decision is made based on a linear programming model taking into account a hotel’s preference for different classes of customers.
Our approach is implemented in five randomly selected hotels of a hotel chain to assess its effectiveness. The difference-in-differences-in-differences estimate with random assignment and matching shows that our approach improves the key operational performance measure–Revenue Per Available Room–by 11.66% with an associated 90% confidence interval of [6.40%, 16.92%], without significantly affecting the average occupancy rate. Since July 1st, 2017, all hotels of the chain have been using our approach for capacity allocation decisions.
Our approach focuses on the revenue improvement rather than revenue optimization, which is aligned with a hotel manager’s incentives. In addition, our recommended solution can be presented in a two-dimension figure, which helps ease a manager’s resistance to a new data-driven approach. Most importantly, our suggested solution is aligned with the long-term goal of the hotel chain, because our approach has been shown to be able to improve hotel revenue without changing the hotel occupancy rate.