Efficient Deep Reinforcement Learning With Imitative Expert Priors for Autonomous Driving.

Area of Science:

• Artificial Intelligence
• Robotics
• Autonomous Systems

Background:

• Deep reinforcement learning (DRL) shows promise for human-like autonomous driving.
• DRL faces challenges with low sample efficiency and complex reward function design.
• These limitations hinder practical applications of DRL in autonomous driving.

Purpose of the Study:

• To propose a novel framework for incorporating human prior knowledge into DRL.
• To enhance sample efficiency and reduce the effort required for reward function design.
• To facilitate the development of practical DRL-enabled autonomous driving systems.

Main Methods:

• Framework integrates expert demonstration, policy derivation, and reinforcement learning.
• Expert behaviors are recorded as state-action pairs for policy derivation.
• Imitative expert policy guides DRL agent via KL divergence regularization.

Main Results:

• Achieved superior performance and significantly improved sample efficiency (60% vs. Soft Actor-Critic).
• Trained agents demonstrated high success rates and human-like driving behaviors.
• Ensemble methods and increased training data further enhanced performance, especially for complex tasks.

Conclusions:

• The proposed framework effectively incorporates human prior knowledge into DRL for autonomous driving.
• It addresses key challenges of sample efficiency and reward function design.
• The method shows strong potential for enabling practical, human-like autonomous driving systems.