ma, zhenyu (Tongji university), cui, yixin (Jilin University), Huang, Yanjun (Tongji University)

From Unsupervised Reinforcement Learning to Continual Reinforcement Learning: Leading Learning from the Relevance to the Whole of Autonomous Driving Decision-Making

Scheduled for presentation during the Invited Session "Data-driven and Learning-based Control Techniques for Intelligent Vehicles" (FrAT1), Friday, September 27, 2024, 10:50−11:10, Salon 1

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

The lifelong learning process in humans, encompassing both cross-domain and within-domain under changing conditions adaptability. It is similarly challenge the realm of artificial intelligence. This issue is particularly pronounced in autonomous driving decision-making, where vehicles must face not only diverse standard environments but also dynamically changing conditions. This paper introduces a novel approach where competence-based unsupervised reinforcement learning (RL) is employed to identify correlations between different policies under real-time varying conditions. Such correlation discovery serves as a foundation for continual learning, further constructing a lifelong learning paradigm. We present detailed theoretical proof of this innovative approach,demonstrating significant improvements over traditional RL baselines.