基于SAC的多智能体深度强化学习算法

doi:10.12263/DZXB.20200243

电子学报 ›› 2021, Vol. 49 ›› Issue (9): 1675-1681.DOI: 10.12263/DZXB.20200243

基于SAC的多智能体深度强化学习算法

肖硕¹^,², 黄珍珍², 张国鹏², 杨树松³, 江海峰², 李天旭²

^1.矿山数字化教育部工程研究中心，江苏徐州 221000
^2.中国矿业大学计算机科学与技术学院，江苏徐州 221000
^3.宁波市轨道交通集团有限公司，浙江宁波 315000

收稿日期:2020-03-09 修回日期:2020-05-14 出版日期:2021-10-21
- 作者简介:
- 肖　硕　男，1981年9月生于江苏徐州。中国矿业大学副教授，硕士生导师，研究方向为人工智能、物联网等。E‑mail:sxiao@cumt.edu.cn
  黄珍珍　女，1981年8月生于江苏徐州。中国矿业大学副研究馆员，研究方向为计算机网络、人工智能、物联网等。E‑mail:huangzhenzhen@cumt.edu.cn
  张国鹏　男，1978年6月生于江苏徐州。中国矿业大学教授，博士生导师，研究方向为人工智能、物联网等。E‑mail:gpzhang@cumt.edu.cn
- 基金资助:
- 国家自然科学基金 (62071470); 徐州市科技计划项目 (KC19011)

Deep Reinforcement Learning Algorithm of Multi⁃agent Based on SAC

XIAO Shuo¹^,², HUANG Zhen-zhen², ZHANG Guo-peng², YANG Shu-song³, JIANG Hai-fei², LI Tian-xu²

^1.Engineering Research Center of Mine Digitalization, Ministry of Education, Xuzhou, Jiangsu 221000, China
^2.School of Computer Sciences and Technology, China University of Mining & Technology, Xuzhou, Jiangsu 221000, China
^3.Operating Branch, Ningbo Rail Transit Group Co., LTD., Ningbo, Zhejiang 315000, China

Received:2020-03-09 Revised:2020-05-14 Online:2021-10-21 Published:2021-09-25
- Supported by:
- National Natural Science Foundation of China (62071470); Xuzhou Science and Technology Project of Jiangsu Province (KC19011)

摘要/Abstract

摘要：

由于多智能体所处环境动态变化，并且单个智能体的决策也会影响其他智能体，这使得单智能体深度强化学习算法难以在多智能体环境中保持稳定. 为了适应多智能体环境，本文利用集中训练和分散执行框架Cen?tralized Training with Decentralized Execution(CTDE)，对单智能体深度强化学习算法Soft Actor?Critic(SAC)进行了改进，引入智能体通信机制，构建Multi?Agent Soft Actor?Critic(MASAC)算法. MASAC中智能体共享观察信息和历史经验，有效减少了环境不稳定性对算法造成的影响.最后，本文在协同以及协同竞争混合的任务中，对MASAC算法性能进行了实验分析，结果表明MASAC相对于SAC在多智能体环境中具有更好的稳定性.

关键词: 多智能体环境, 集中训练, 分散执行, 多智能体深度强化学习

Abstract:

Due to the dynamic change of multi-agent environment, and the decision of single agent will affect other agents, it is difficult for the deep reinforcement learning algorithm of single agent to maintain stability in multi-agent environment. In order to adapt to multi-agent environment, this paper uses centralized training and decentralized execution framework (CTDE) to improve single agent deep reinforcement learning algorithm soft actor-critic (SAC). By introducing agent communication mechanism, in multi-agent soft actor-critic (MASAC), agents share observation information and historical experience, which effectively reduces the impact of environmental instability on the algorithm. Finally, in the task of cooperation and cooperation and competition, the performance of MASAC algorithm is analyzed experimentally. The results show that MASAC has better stability than SAC in multi-agent environment.

Key words: multi-agent environments, centralized training, decentralized execution, multi-agent deep reinforcement learning

中图分类号:

TP391

肖硕, 黄珍珍, 张国鹏, 等. 基于SAC的多智能体深度强化学习算法[J]. 电子学报, 2021, 49(9): 1675-1681.

Shuo XIAO, Zhen-zhen HUANG, Guo-peng ZHANG, et al. Deep Reinforcement Learning Algorithm of Multi⁃agent Based on SAC[J]. Acta Electronica Sinica, 2021, 49(9): 1675-1681.

图/表 8

图1 MASAC网络模型

图2 MASAC中的通信模型

图3 写操作模型

图4 CC环境不同算法获得的奖励

表1 智能体到达目标的概率和距离目标的平均距离

算法	到达目标的概率	平均距离
MASAC	87.5%	0.126
SAC	19.3%	0.437

图5 physical deception环境示意图

图6 physical deception训练20000轮效果图

图7 physical deception训练60000轮效果图

参考文献 16

1	SilverD, HuangA, MaddisonC J, et al. Mastering the game of Go with deep neural networks and tree search[J]. Nature, 2016, 529(7587): 484 - 489.
2	SilverD, SchrittwieserJ, SimonyanK, et al. Mastering the game of Go without human knowledge[J]. Nature, 2017, 550(7676): 354 - 359.
3	周沛, 陈后金, 于泽宽, 等. 跨模态医学图像预测综述[J]. 电子学报, 2019, 47(1):220 - 226.
	ZhouP, ChenH J, YuZ K, et al. A review of multimodal medical image prediction [J]. Acta Electronica,Sinica,2019, 47 (1): 220 - 226. (in Chinese)
4	LoweR, FoersterJ, BoureauY L, et al. On the pitfalls of measuring emergent communication[A]. Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems[C]. Montreal: ICAAMS, 2019. 693 - 701.
5	WangX, ChenW, WuJ, et al. Video captioning via hierarchical reinforcement learning[A]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition[C]. Hawaii: CVPR, 2018. 4213 - 4222.
6	郑兴华, 孙喜庆, 吕嘉欣,等. 基于深度学习和智能规划的行为识别[J]. 电子学报, 2019, 47(8):1661 - 1668.
	ZhengX H, SunX Q, LVJ X, et al. Behavior recognition based on deep learning and intelligent planning [J]. Acta Electronica Sinica, 2019, 47 (8): 1661 - 1668. (in Chinese)
7	SchulmanJ, LevineS, AbbeelP, et al. Trust region policy optimization[A]. International Conference on Machine Learning[C]. Lille: ICML, 2015. 1889 - 1897.
8	闻佳, 王宏君, 邓佳, 等. 基于深度学习的异常事件检测[J]. 电子学报, 2020,48(2):308 - 313.
	WenJ, WangH J, DengJ, et al. Abnormal event detection based on deep learning [J]. Acta Electronica Sinica, 2020, 48 (2): 308 - 313. (in Chinese)
9	AbdallahS, KaisersM. Addressing environment non-stationarity by repeating Q-learning updates[J]. The Journal of Machine Learning Research, 2016, 17(1): 1582 - 1612.
10	FoersterJ N, FarquharG, AfourasT, et al. Counterfactual multi-agent policy gradients[A]. Thirty-second AAAI Conference on Artificial Intelligence[C]. New Orleans: AAAI, 2018. 2974 - 2982.
11	LoweR, WuY, TamarA, et al. Multi-agent actor-critic for mixed cooperative-competitive environments[A]. Advances in Neural Information Processing Systems[C]. Long Beach: NIPS, 2017. 6379 - 6390.
12	HaarnojaT, ZhouA, AbbeelP, et al. Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor[A]. International Conference on Machine Learning[C]. Stockholm: ICML, 2018. 1856 - 1865.
13	HaarnojaT, TangH, AbbeelP, et al. Reinforcement learning with deep energy-based policies[A]. Proceedings of the 34th International Conference on Machine Learning[C]. Sydney: ICML, 2017. 1352 - 1361.
14	DasA, KotturS, MouraJ M F, et al. Learning cooperative visual dialog agents with deep reinforcement learning[A]. Proceedings of the IEEE International Conference on Computer Vision[C]. Venice: ICCV, 2017. 2951 - 2960.
15	曹源,唐涛,徐田华,穆建成.形式化方法在列车运行控制系统中的应用[J].交通运输工程学报, 2010, 10(1):112 - 126.
	CaoYuan, TangTao, XuTianhua, MuJiancheng. Application of formal method in train operation control system [J]. Journal of Transportation Engineering, 2010, 10 (1): 112 - 126. (in Chinese)
16	吴胜权,黄振晖,曹源. 有轨电车路权配置与信号系统选择[J]. 中国铁路, 2014, (8):97 - 99.
	WuShengquan, HuangZhenhui, CaoYuan. Tram right of way configuration and signal system selection [J]. China Railway, 2014, (8): 97 - 99. (in Chinese)

基于SAC的多智能体深度强化学习算法

Deep Reinforcement Learning Algorithm of Multi⁃agent Based on SAC

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 16

相关文章 1

编辑推荐

Metrics