Optimizing hyperparameters of deep reinforcement learning for autonomous driving based on whale optimization...

Area of Science:

• Artificial Intelligence
• Robotics
• Control Systems

Background:

• Deep Reinforcement Learning (DRL) agents learn optimal policies through reward functions without prior environmental knowledge.
• Hyperparameter tuning is critical for DRL efficiency and performance, presenting a significant challenge.
• Autonomous driving systems require robust control strategies capable of handling complex, continuous action spaces.

Purpose of the Study:

• To optimize hyperparameters for the Deep Deterministic Policy Gradient (DDPG) algorithm using a swarm-based approach.
• To enhance the control strategy of DRL agents in autonomous driving scenarios.
• To address the challenge of accurate hyperparameter estimation in DRL training.

Main Methods:

• Employed the Whale Optimization Algorithm (WOA), a swarm-based metaheuristic, for hyperparameter optimization.
• Applied the optimized DDPG algorithm to an autonomous driving control problem within the TORCS simulation environment.
• Compared the performance of the DDPG agent with optimized hyperparameters against one with reference hyperparameters.

Main Results:

• Hyperparameter optimization using WOA led to maximized total rewards for the DDPG agent.
• The optimized DDPG agent demonstrated improved performance across testing episodes.
• A more stable driving policy was achieved with the optimized DDPG hyperparameters.

Conclusions:

• Swarm-based optimization, specifically WOA, is effective for tuning DRL hyperparameters in autonomous driving.
• Optimized DDPG hyperparameters significantly enhance learning efficiency and policy stability.
• The proposed method offers a viable solution for improving DRL performance in complex control tasks.