Dynamic Pricing On Air Tickets Based On Reinforcement Learning

With the rapid development of the Internet and tourism industry,online travel agencies(OTAs)have gradually replaced traditional travel agents with excellent services and become a necessary part when people are going to travel and buy tickets.Among the many businesses operated by OTAs,air ticket business is rather important.Considering their own profit,OTAs hope to maximize the total profit for selling air tickets within a period of time by dynamically adjusting the price of air ticket products.Currently,most OTAs earn profits by selling tickets for airlines and charging an additional commission based on this.Many OTAs try to adjust commissions based on their industry experience to obtain more profits.However,the demand for air tickets and the behavior patterns of users are very complicated in reality.Therefore,the application of methods such as expert experience and mathematical models has many disadvantages in price adjustment decisions.Based on the characteristics of reinforcement learning algorithms that are good at solving strategic optimization problems,this paper uses reinforcement learning algorithms to study dynamic pricing strategies.On the other hand,interaction with the environment is a necessary process during the training of reinforcement learning.Besides,in the dynamic pricing problem,directly interacting with the real environment will inevitably bring a huge negative effect on the market.Considering these two factors,how to simulate the behavior of users and develop a simulation environment that can reflect the market response based on this is a very important research.Therefore,we proposed a simulation model for ticket purchase.This model can generate a group of users with different features based on historical data,and then simulate the users’ purchase behavior based on other information such as ticket prices.The model can also continuously adjust its parameters based on historical data and newly generated data,it has the ability of self-learning which makes the simulation model perform better.At the same time,the performance of reinforcement learning algorithms will be affected by the unpredictable state due to the high uncertainty of the market and users in the pricing problem.At the same time,the user’s consumer demand will also change as external conditions change.Using the same pricing model cannot take into account users with different needs,which will inevitably lead to the effect of the model not being good enough.So based on these points,we propose a regime-switching recurrent neural network reinforcement learning algorithm.Finally,the experimental results on real-world data sets show that our user simulation model can simulate user feedback well,and our reinforcement learning dynamic pricing algorithm performs well in comparison with other algorithms.