Deep Reinforcement Learning on Autonomous Driving Policy With Auxiliary Critic Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Deep reinforcement learning (DRL) offers potential for complex decision-making, particularly in autonomous driving.
• Autonomous driving systems require robust policies to navigate diverse and dynamic traffic scenarios.
• Existing DRL methods face challenges in exploration efficiency and potential overestimation bias.

Purpose of the Study:

• To propose an end-to-end autonomous driving policy learning method using DRL.
• To enhance exploration efficiency and control performance in autonomous driving tasks.
• To mitigate overestimation bias in DRL training for autonomous vehicles.

Main Methods:

• Implemented a DRL approach based on Proximal Policy Optimization (PPO).
• Integrated a curiosity-driven method using Recurrent Neural Networks (RNN) for intrinsic reward generation.
• Introduced an auxiliary dual critic network to reduce overestimation bias during training.

Main Results:

• The proposed method demonstrated improved training efficiency compared to other DRL techniques.
• Enhanced control performance was observed in lane-keeping and overtaking tasks within the TORCS simulator.
• The curiosity-driven exploration and dual critic network effectively addressed exploration and bias issues.

Conclusions:

• The novel DRL method significantly improves autonomous driving policy learning.
• The integration of curiosity-driven exploration and dual critics enhances both training and performance.
• This approach shows promise for developing more capable and efficient autonomous driving systems.