Neural Network Architecture Optimization Based on Virtual Reward Reinforcement Learning

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/84082

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/84082

Title:	Neural Network Architecture Optimization Based on Virtual Reward Reinforcement Learning
Authors:	江玟萱;Chiang, Wen-Hsuan
Contributors:	資訊管理學系
Keywords:	神經架構搜索;強化學習;近端策略優化;神經網絡優化;機器學習;Neural Architecture Search;Reinforcement Learning;Proximal Policy Optimization;Neural Network Optimization;Machine Learning
Date:	2020-08-20
Issue Date:	2020-09-02 18:02:37 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來機器學習越來越受大眾歡迎，造成越來越多學者、業者、工程師等都進行相關的研究與應用。只要他們對於資料不夠理解，就有可能造成資訊的誤解或是模型的偏差，因為他們抓取的特徵就是一個機器學習的指標。為了避免手動抓取特徵的上述狀況，我們可以透過機器建立神經網路。我們的研究使用預測器來構立虛擬地圖。使用此虛擬地圖來訓練代理人，讓它可以找到良好的神經網絡體結構。但是獎勵函數有一些改變，因此我們在本研究中提出了四種模型。在實驗過程中，我們分析了四種模型的每個參數的實驗結果。並意識到模型穩定性的重要性。如果模型不穩定，則獲得的正確率的差距可能太大。然而我們的模型在正確率以及穩定性方面具有良好的性能。;Abstract-- In recent years, machine learning has become more and more popular, causing more and more scholars, practitioners, and engineers to conduct related research and applications. If they don′t understand the data well, it may cause misunderstanding of the information or deviation of the model, because the feature they capture is an indicator of machine learning. In order to avoid the above situation of manually grabbing features, we can build neural networks through machines. Our research uses a predictor to build a virtual map. Using this virtual map to train agents to find the good neural network architecture. But the reward function has some changes, so we proposed four models in this research. During the experiment, we analyze the experimental results of each parameter for the four models. And realize the importance of the model stability. If the model is unstable, the gap between the obtained accuracy may be too large. However, our model has a good performance in accuracy and stability.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	246	View/Open

社群 sharing

Loading...