Adaptive Adversarial Training for Meta Reinforcement Learning

被引：2

作者：

Chen, Shiqi ^{[1
,3
]}

Chen, Zhengyu ^{[1
,2
]}

Wang, Donglin ^{[1
,2
]}

机构：

[1] Westlake Univ, Sch Engn, AI Div, Machine Intelligence Lab MiLAB, Hangzhou, Peoples R China

[2] Westlake Inst Adv Study, Inst Adv Technol, Hangzhou, Peoples R China

[3] Nanyang Technol Univ, Wee Kim Wee Sch Commun & Informat, Singapore, Singapore

来源：

2021 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN) | 2021年

关键词：

Adversarial Training; Meta Reinforcement Learning; GAN; Robustness;

D O I：

10.1109/IJCNN52387.2021.9534316

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

Meta Reinforcement Learning (MRL) enables an agent to learn from a limited number of past trajectories and extrapolate to a new task. In this paper, we attempt to improve the robustness of MRL. We build upon model-agnostic meta-learning (MAML) and propose a novel method to generate adversarial samples for MRL by using Generative Adversarial Network (GAN). That allows us to enhance the robustness of MRL to adversal attacks by leveraging these attacks during meta training process.

引用

页数：8

共 28 条

[1]

Carlini N, 2017, PROCEEDINGS OF THE 10TH ACM WORKSHOP ON ARTIFICIAL INTELLIGENCE AND SECURITY, AISEC 2017, P3, DOI 10.1145/3128572.3140444

[2]

Chen Z, 2021, 2021 INT C AC SPEECH

[3]

Doshi-Velez Finale, 2016, IJCAI (U S), V2016, P1432

[4]

Duan Y., 2016, Rl2: Fast reinforcement learning via slow reinforcement learning

[5]

Finn C, 2017, PR MACH LEARN RES, V70

[6]

Goldblum M., 2019, ARXIV191000982

[7]

Goodfellow Ian J., 2014, INT C LEARNING REPRE

[8]

Huang Sandy, 2017, P ICLR

[9]

Kos D., 2017, Delving into adversarial attacks on deep policies

[10]

Lee H B, 2019, ARXIV190512914

← 1 2 3 →