Please wait a minute...
 
主管单位:广东省科学技术厅
主办单位:广东省科技合作研究促进中心
编辑出版:《电脑与电信》编辑部
ISSN 1008-6609 CN 44-1606/TN
邮发代号:46-95
国内发行:广东省报刊发行局
《电脑与电信》唯一官方网站。
电脑与电信  2023, Vol. 1 Issue (5): 38-    DOI: 10.15966/j.cnki.dnydx.2023.05.020
  基金项目 本期目录 | 过刊浏览 | 高级检索 |
模仿学习算法的研究与实现
南京理工大学紫金学院 计算机学院
Research and Implementation of Imitation Learning Algorithm
Nanjing University of Science and Technology Zijin College
全文: PDF( KB)  
输出: BibTeX | EndNote (RIS)      
摘要 
为优化强化学习因为奖励函数不明确造成极大误差的弊端,深入研究并实现了模仿学习算法中的行为克隆算法 和数据聚合算法。通过活动图对算法流程建模,通过类图对各类之间的关系建模,通过顺序图对核心交互流程建模。根据实 验结果,比较行为克隆算法和数据聚合算法的优缺点,发现行为克隆算法可以离线训练避免真实环境的交互,但会使错误累加 导致结果误差;数据聚合算法必须在线与环境交互,根据策略选择观测值对应状态,解决误差累积的问题。
服务
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章
关键词 强化学习模仿学习行为克隆算法数据聚合算法    
Abstract:In order to optimize reinforcement learning for the great errors causing by the unclear reward function, this paper deeply studies and implements the behavior cloning algorithm and data aggregation algorithm in the imitation learning algorithm. The algorithm flow is modeled by activity diagram, the relationship between classes is modeled by class diagram, and the core interaction process is modeled by sequence diagram. According to the experimental results, this paper compares the advantages and disadvantages of the behavior cloning algorithm and the data aggregation algorithm, and discovers that the behavior cloning algorithm offline training can avoid interaction with the real environment, but error accumulation will lead to error results; data aggregation algorithms must interact with the environment online, and select the corresponding state of the observation value according to the strategy to solve the problem of error accumulation.
Key wordsreinforcement learning    imitation learning    behavior cloning algorithm    data aggregation algorithm
年卷期日期: 2023-05-10      出版日期: 2024-01-24
引用本文:   
张羽萌 季晓君. 模仿学习算法的研究与实现[J]. 电脑与电信, 2023, 1(5): 38-.
ZHANG Yu-meng JI Xiao-jun. Research and Implementation of Imitation Learning Algorithm. Computer & Telecommunication, 2023, 1(5): 38-.
链接本文:  
https://www.computertelecom.com.cn/CN/10.15966/j.cnki.dnydx.2023.05.020  或          https://www.computertelecom.com.cn/CN/Y2023/V1/I5/38
No related articles found!
No Suggested Reading articles found!
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
  Copyright © 电脑与电信 All Rights Reserved.
地址:广州市连新路171号广东国际科技中心 邮编:510033
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn
粤ICP备05080322号-4