Safe Decision Controller for Autonomous DrivingBased on Deep Reinforcement Learning inNondeterministic Environment.

Area of Science:

• Cyber-physical systems
• Artificial Intelligence
• Autonomous Driving

Background:

• Deep reinforcement learning (DRL) is widely used for autonomous driving decision-making.
• Black-box DRL lacks safety guarantees and interpretability in uncertain, complex environments.
• Uncontrolled uncertainties and reward function settings pose significant challenges.

Purpose of the Study:

• To propose a formal security reinforcement learning (RL) method for autonomous driving systems.
• To enhance the safety and interpretability of RL-based decision-making.
• To address challenges posed by complex environments and uncertainties.

Main Methods:

• Developed an environmental modeling approach to quantify nondeterministic factors.
• Formalized reward machine structure using the environment model for RL reward setting.
• Generated a control barrier function to ensure safer state behavior policies.

Main Results:

• The proposed method effectively models environmental uncertainties.
• Formalized reward machines improve reward-function setting in RL.
• Control barrier functions enhance policy safety in autonomous driving.

Conclusions:

• The formal security reinforcement learning method improves safety and interpretability in autonomous driving.
• The approach is verified effective in intelligent driving scenarios like overtaking and lane-changing.
• This work contributes to safer and more reliable autonomous systems.