Abstract: In order to solve the problem that the self-driving mining truck is easy to lose the ability to adapt to the previous
strategy in the complex environment of the mine,a deep reinforcement learning lane keeping control algorithm considering
sample retraining is proposed. Firstly,by considering the characteristics of the target network update parameters,a periodic experience
extraction retraining model is derived,and the retraining round interval is incorporated into the traditional target network
update parameter model. Then,in order to avoid the influence of noise on the model,the experience playback buffer is set
in a smaller sampling range. The influence of noise and unrelated experience on the model will be reduced,and the system robustness
under extreme operating conditions will be enhanced. Finally,considering the typical cross-shaped road of open-pit
mine,the vehicle position is set at the crossroads in CARLA,and the average reward obtained under the fixed number of rounds
is used as the key performance index of the simulation. The experimental results show that the proposed periodic retraining deep
Q network (PR-DQN) strategy effectively reduces the fluctuation in the training process,makes the model converge faster,effectively
improves the performance of the model in non-stationary environment tasks,and shows significant advantages in stability
and generalization ability.