Path planning of mobile robot based on improved double deep Q-network algorithm.

Area of Science:

• Robotics and Artificial Intelligence
• Machine Learning
• Path Planning Algorithms

Background:

• Traditional deep reinforcement learning algorithms face challenges with slow network convergence, unstable reward convergence, and inefficient path planning.
• Existing methods often struggle in complex environments, leading to suboptimal path lengths and increased planning times.

Purpose of the Study:

• To propose an advanced path planning algorithm, BiLSTM-D3QN (Bidirectional Long and Short-Term Memory Dueling Double Deep Q-Network), that addresses the limitations of traditional deep reinforcement learning.
• To enhance network convergence speed, reward convergence stability, and overall path planning efficiency.

Main Methods:

• Integration of Bidirectional Long Short-Term Memory (BiLSTM) networks for improved memory and decision-making stability.
• Incorporation of Dueling Network architecture to mitigate Q-value overestimation and accelerate network updates.
• Implementation of Adaptive Experience Replay with a frequency penalty function for efficient data extraction.
• Introduction of an adaptive action selection mechanism to optimize exploration.

Main Results:

• BiLSTM-D3QN demonstrates superior network convergence speed, planning efficiency, reward convergence stability, and success rates in simple environments compared to traditional Deep Reinforcement Learning.
• In complex environments, BiLSTM-D3QN achieved a 20m shorter path length, 7 fewer turning points, 0.54s faster planning time, and a 10.4% higher success rate than the improved ERDDQN algorithm.

Conclusions:

• The proposed BiLSTM-D3QN algorithm significantly outperforms existing methods in both convergence speed and path planning performance.
• BiLSTM-D3QN offers a more stable and efficient solution for complex path planning tasks in reinforcement learning.