基于策略学习的机票动态定价算法

机票动态定价旨在构建机票售价策略以最大化航班座位收益.现有机票定价算法都建立在提前预测各票价等级的需求量基础之上,会因票价等级需求量的预测偏差而降低模型性能.为此,提出基于策略学习的机票动态定价算法,其核心是不再预测各票价等级的需求量,而是将机票动态定价问题建模为离线强化学习问题.通过设计定价策略评估和策略更新的方式,从历史购票数据上学习具有最大期望收益的机票动态定价策略.同时设计了与现行定价策略和需求量预测方法的对比方法及评价指标.在两趟航班的多组定价结果表明:相比于现行机票销售策略,策略学习算法在座位收益上的提升率分别为30.94％和39.96％,且比基于需求量预测方法提升了6.04％和3.36％.

Approach for Dynamic Flight Pricing Based on Strategy Learning

The core of the dynamic flight pricing is to yield a pricing strategy with maximum seat revenue. The state-of-the-art flight pricing approaches are built on forecasting the fare demand. They suffer low profit due to the inaccurate prediction. To tackle the above issue, an approach for dynamic flight pricing based on strategy learning is proposed. That approach resorts to reinforcement learning to output pricing strategy with the highest expected return. That strategy is learned by iteratively policy evaluation and policy improvement. The rate of profit improvement on the two flights is empirically 30.94% and 39.96% over the existing pricing strategy, while that rate is 6.04% and 3.36% over the demand forecasting algorithm.