Adversarial imitation learning based on a high sample efficiency framework

View at publisher Summarise with VIRAT
Authors: Quan Liu, Zenghui Liu, Jiahui Li
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper proposes a high-sample-efficiency off-policy deep reinforcement learning framework based on Discriminator- Actor-Critic. An efficient experience replay mechanism is developed by incorporating data augmentation and a stochastic soft dynamic buffer. The buffer is designed to evict samples with large deviations from the target policy, balancing policy inconsistency while avoiding catastrophic forgetting. In addition, the Wasserstein distance replaces the Jensen–Shannon divergence to enhance training stability and serve as a reliable indicator of training progress, and the Soft Actor-Critic algorithm is adopted to improve agent exploration. Experiments show that the proposed method achieves faster and more stable convergence than baseline approaches, and maintains competitive performance even when expert trajectories are limited.

Fields of Study & Subjects

Reinforcement Learning in Robotics Domain Adaptation and Few-Shot Learning Adaptive Dynamic Programming Control

Publication details

Journal / source
International Conference on Optics, Computer Applications, and Materials Science (CMSD)
Pages
59–59
Publication date
1 April 2026
Language
EN