請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92127
標題: | 長短期記憶神經網路之消費者偏好學習與動態訂價運用 Adopting Long Short-Term Memory Neural Network for Consumer Preference Learning and Dynamic Pricing |
作者: | 王文謙 Wen-Chian Wang |
指導教授: | 吳政鴻 Cheng-Hung Wu |
關鍵字: | 動態規劃,動態訂價,需求學習,深度學習,長短期記憶神經網路, Dynamic programming,Dynamic pricing,Demand learning,Deep learning,LSTM neural network, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 需求學習(Demand Learning)可以幫助供應商了解消費者喜好來達到滿足市場需求,但往往受限於歷史資料不足及環境變動快速等問題,無法得到最佳的效果。因應此問題,朱婉琳(2019)提出一種結合動態規劃(Dynamic Programming)與深度學習(Deep Learning)之技術應用在動態訂價(Dynamic Pricing)與需求學習中,該模型使用長短期記憶神經網路(Long Short-Term Memory neuron network, LSTM)對各種市場環境進行消費者需求學習,在面對新環境時透過歷史資料估計新環境下消費者支付意願(Willing to Pay, WTP)之參數,透過動態訂價求解出最佳決策並逼近最佳利潤。
然而,朱婉琳之研究僅透過神經網路學習消費者支付意願分布之平均值,並假設支付意願服從標準差為固定值之常態分布(Normal Distribution),並不符合實際市場中的狀況:消費者支付意願為服從任意常態分布,即常態分布之平均值及標準差值皆是變動的,亦或是支付意願服從其他各種不同分布。因此,本研究將使用神經網路模型學習消費者支付意願分布,其中透過KL Divergence 作為神經網路的損失函數(Loss)去學習常態分布之平均值及標準差,或是學習伽瑪分布的參數,更符合實際市場銷售情形,並透過動態訂價求解,以達到供應商販售時效性商品之最大化利潤。 最後透過模擬訂價驗證在各式各樣的市場環境下,本研究提出的模型僅需要用很少的成本學習消費者的需求,且模擬訂價的結果與動態規劃求解最佳利潤的差異很少,結果證實,本研究之模型能在未知的環境下擁有良好的能力。 Demand Learning can help suppliers understand consumer preferences to meet market demand. However, it is often limited by insufficient historical data and rapid environmental changes, and cannot achieve the best results. In response to this problem, Zhu(2019) proposed a technology that combines dynamic programming and deep learning to be used in dynamic pricing and demand learning. This model uses Long Short-Term Memory neuron network(LSTM) conducts consumer demand learning in various market environments, and uses historical data to predict consumer willingness to pay (WTP) in the new environment through historical data. Dynamic pricing solves the best decision and approaches the best profit. However, Zhu''s research only used neural networks to learn the average value of consumers'' WTP distribution, and assumed that WTP obeys a normal distribution with a fixed standard deviation, which is not consistent with the situation in the actual market: consumers The WTP is subject to any normal distribution, that is, the mean and standard deviation of the normal distribution are changing, or the WTP is subject to various other different distributions. Therefore, this study will use a neural network model to learn the distribution of consumers'' willingness to pay, using KL Divergence as the loss function of the neural network to learn the mean and standard deviation of the normal distribution, or learn the gamma distribution The parameters are more in line with the actual market sales situation and are solved through dynamic pricing to maximize the supplier''s profit from selling time-sensitive goods. Finally, it is verified through simulated pricing that in various market environments, the model proposed in this study only needs to learn consumer needs at a very small cost, and the results of simulated pricing are different from those of dynamic programming to solve the optimal profit. Very rarely, the results confirm that the model in this study has good capabilities in unknown environments. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92127 |
DOI: | 10.6342/NTU202400728 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 工業工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf | 4.03 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。