NeuroAction: a neuroevolutionary approach to reinforcement learning for autonomous vehicles

Area of Science:

• Artificial Intelligence
• Robotics
• Control Systems

Background:

• Deep Reinforcement Learning (DRL) for autonomous vehicles typically uses decoupled perception-action protocols.
• A key limitation is the reliance on backpropagation, which requires aggregating multiple objectives into a single loss function.
• This aggregation restricts personalized driving behaviors and user-defined preferences.

Purpose of the Study:

• To introduce NeuroAction, a multi-objective neuroevolutionary method for reinforcement learning-based autonomous driving.
• To address the challenge of optimizing multiple, potentially conflicting, objectives simultaneously.
• To enable the generation of diverse driving policies catering to user-specific preferences.

Main Methods:

• Formulating autonomous vehicle control as a multiobjective optimization problem.
• Utilizing multiobjective evolutionary algorithms (EMO) to solve the optimization problem.
• Generating a Pareto front of optimal policy networks representing different trade-offs.

Main Results:

• The proposed NeuroAction framework was investigated on a benchmark DRL-based autonomous driving task.
• Performance evolution was analyzed using three different EMO algorithms.
• The method demonstrated the capability to optimize multiple objectives concurrently for autonomous driving policies.

Conclusions:

• NeuroAction offers a flexible framework for multi-objective reinforcement learning in autonomous driving.
• The approach allows for the creation of driving policies that align with diverse user preferences.
• This method overcomes the limitations of single-objective optimization in DRL for autonomous systems.