搭配優先經驗回放與經驗分類於深度強化式學習之記憶體減量技術;Memory Reduction through Experience Classification with Prioritized Experience Replay in Deep Reinforcement Learning

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/81372

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/81372

題名:	搭配優先經驗回放與經驗分類於深度強化式學習之記憶體減量技術;Memory Reduction through Experience Classification with Prioritized Experience Replay in Deep Reinforcement Learning
作者:	沈楷桓;Shen, Kai-Huan
貢獻者:	電機工程學系
關鍵詞:	深度強化式學習;深度確定性決策梯度;優先經驗回放;Deep reinforcement learning;Deep deterministic policy gradient;Prioritized experience replay
日期:	2019-06-25
上傳時間:	2019-09-03 15:49:01 (UTC+8)
出版者:	國立中央大學
摘要:	優先經驗回放在許多強化式線上學習演算法中被廣泛的使用，使過往經驗能夠被更有效率的利用，但在其中使用大量記憶體顯著的消耗了系統存儲。因此，在這份研究中我們提出了一個分區及分類機制減輕上述的效應。由計算學習歷程中所得經驗的時差誤差(temporal-difference error)可以得知若某些情況尚未學會，將導致其誤差較大。而我們利用這個資訊設計出多個區段來控管記憶體複寫經驗：將上述經驗的時差誤差排序組成累積分布函數，並導入一個新的超參數(hyper-parameter S)控制區段數量，將累積分布函數分成S個區段且算出各區段的成員數量與其邊界；新的經驗將根據其時差誤差大小與所屬區段中的經驗的記憶體位址交換資料來改變其存活時間；另訂出一隨記憶體複寫指標(write-data pointer)移動之凍結區(frozen region)，所有歸於此區記憶體位址的經驗將不會被置換，以免經驗透過交換存活過久；隨著網路更新，此分區資訊將週期性更新，避免使用過時資訊控制記憶體。透過這個機制改變已學會經驗與延長高價值經驗的存活時間；而上述分區及分類機制由時間差分誤差的資訊計算取得。所提出的機制結合深度確定性決策梯度演算法於倒立擺和倒立雙擺問題驗證。從實驗可知，我們所提出的機制有效的減少記憶體中冗餘內容，降低記憶體內資料的相關性並以額外的時間差分誤差計算為代價達到更佳的學習效能和更少的記憶體使用量。;Prioritized experience replay has been widely used in many online reinforcement learning algorithms, providing high efficiency in exploiting past experiences. However, a large replay buffer consumes system storage significantly. Thus, in this paper, a segmentation and classification scheme is proposed to change the lifetime of the well-learned experiences and the valuable experiences. The segmentation and classification are achieved by using the information of temporal-difference errors (TD errors). With the ranking of TD errors, we further calculate the cumulative density function and separate it into S segments, where S is a newly introduced hyper-parameter; then, obtain the information of each segment. The incoming new experiences will be classified to the corresponding segments according to their TD error and will be swapped with the data in the same segment. Besides, we define a frozen region, which follows the write-data pointer of the replay buffer, to avoid experiences living too long with the proposed scheme. The proposed scheme is incorporated in the Deep Deterministic Policy Gradient (DDPG) algorithm, and the Inverted Pendulum and Inverted Double Pendulum tasks are used for verification. From the experiments, our proposed mechanism can effectively remove the buffer redundancy. Thus, better learning performance with reduced memory size is achieved at the cost of additional computations of updated TD errors.
顯示於類別:	[電機工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	119	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....