Noisy Dueling Double Deep Q-Network algorithm for autonomous underwater vehicle path planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Oceanography

Background:

• Autonomous Underwater Vehicle (AUV) path planning in complex ocean currents is challenging.
• Traditional reinforcement learning methods struggle with environmental exploration and generalization.
• Existing algorithms like RRT*, DQN, and D3QN have limitations in dynamic environments.

Purpose of the Study:

• To develop a novel algorithm for improved AUV path planning.
• To enhance the success rate and reduce travel time of AUV missions.
• To address the generalization limitations of traditional reinforcement learning in AUV navigation.

Main Methods:

• Proposed the Noisy Dueling Double Deep Q-Network (ND3QN) algorithm.
• Modified the reward function and incorporated a noisy network into the D3QN framework.
• Conducted simulation experiments in realistic ocean terrain and current conditions.

Main Results:

• The ND3QN algorithm demonstrated a higher success rate for AUV path planning.
• ND3QN achieved significantly shorter travel times compared to classical algorithms.
• The algorithm generated smoother paths, indicating improved navigation efficiency.

Conclusions:

• The ND3QN algorithm offers a superior approach to AUV path planning in challenging marine environments.
• The proposed method effectively overcomes the exploration and generalization issues of traditional reinforcement learning.
• ND3QN provides a robust and efficient solution for practical AUV applications.