Research on Vehicle Obstacle Avoidance Navigation Algorithm Based on Neural Ordinary Differential Equation Deep Q-Network
The effective obstacle avoidance of vehicles is crucial for enhancing the safety and reliability of autono-mous driving systems.In recent years,deep reinforcement learning methods represented by Deep Q-Network(DQN)algo-rithms have been widely applied to research on vehicle obstacle avoidance.However,the multilayer perceptrons in the DQN algorithm tend to overfit when handling small sample data,leading to insufficient generalization ability.To address this is-sue,a Deep Q-Network algorithm based on neural ordinary differential equation is proposed,referred to as NODQN.The proposed NODQN algorithm replaces the multilayer perceptrons in the traditional DQN algorithm with Neural ODE for Q-value function approximation.Experimental results demonstrate that in three typical simulation scenarios,the NODQN algo-rithm outperforms the DQN algorithm across key performance metrics,including success rate,collision rate and average re-ward.Additionally,the NODQN algorithm exhibits a lower average maximum offset compared to the DQN algorithm.In conclusion,the NODQN algorithm effectively improves the overfitting issue of the traditional DQN algorithms when handling small sample data,demonstrating stronger generalization ability and higher stability,thereby enhancing the safety and relia-bility of the system in vehicle obstacle avoidance tasks.
NETL
NSTL
万方数据
