基于强化学习的机票买卖智能体研究
首发时间:2020-07-07
摘要:针对当前民航业机票销售领域存在的:价格敏感型乘客的出行需求、机票销量和客机落座率不高的问题,本文构建了一种新型交易模式--在不影响市场大环境的前提下,由代理商为部分乘客代购机票,通过在低价时购票获取利润。 由于民航业务特性,机票价格频繁波动,人工智能可以从变动的价格中找出人类很难发现的一些规律。强化学习在解决序贯决策问题上,表现出突出的效果,因而受到广泛关注和应用。机票买卖从某种角度上可以看作一种序贯决策问题,本文较具创新性地尝试利用强化学习中的DQN,结合CNN卷积神经网络,提出了一种机票买卖Agent模型,并基于Keras框架实现了该算法模型。该模型针对代购订单,选择最佳时刻进行代购,目标是实现已有订单的收益最大化。本文通过实验结果证明了此模型在处理这一问题的好处。
关键词: 强化学习 DQN CNN卷积神经网络 智能决策 机票交易
For information in English, please click here
Research on Air Ticket Trading Agent Based on Reinforcement Learning
Abstract:In view of the current problems in the air ticket sales field of the civil aviation industry: the travel needs of price-sensitive passengers, the sales volume of air tickets and the low seating rate of passenger aircraft, this paper builds a new transaction model-without affecting the market environment, by The agent purchases air tickets for some passengers and obtains profits by purchasing tickets at low prices. Due to the characteristics of civil aviation business, ticket prices fluctuate frequently, and artificial intelligence can find some laws that are difficult for humans to find out from the changed prices. Reinforcement learning has shown outstanding results in solving sequential decision-making problems, so it has been widely concerned and applied. Air ticket sales can be regarded as a sequential decision problem from a certain angle. This paper attempts to use DQN in reinforcement learning in a more innovative way, combined with CNN convolutional neural network, and proposes an air ticket sales agent model, which is based on Keras. The framework implements the algorithm model. The model selects the best moment for purchasing orders for purchasing orders. The goal is to maximize the revenue of existing orders. This paper proves the benefits of this model in dealing with this problem through experimental results.
Keywords: Reinforcement learning DQN CNN convolutional neural network intelligent decision-making ticket transaction
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于强化学习的机票买卖智能体研究
评论
全部评论0/1000