Training method and device for unmanned vehicle control model

Abstract

The invention discloses a training method and device for an unmanned vehicle control model. Aiming at each moment during model training; the method comprises the steps of determining a feature matrixcomposed of historical environment features used for calculating awards at the last moment and current environment features determined according to current environment information; then, based on theimportance degree of the current environment feature and each historical environment feature on the feature matrix, selecting features used for calculating rewards at the current moment from the feature matrix, determining the rewards according to the current environment features and the selected features to train the unmanned vehicle control model, and performing unmanned vehicle control according to the trained model after the training is finished. Because the calculation of the features of the rewards is determined based on the overall importance degree of the feature pairs of the feature pairs including the historical environment features, more effective rewards can be determined based on the change of the environment information during training, the problem of sparse rewards is solved, and the cost and time are saved.

Description

technical field [0001] The present application relates to the field of unmanned driving technology, in particular to a training method and device for an unmanned vehicle control model. Background technique [0002] At present, the main problem that needs to be solved in the control method of unmanned vehicles in the field of unmanned driving technology is how to avoid obstacles for unmanned vehicles. Usually, the obstacle avoidance process of unmanned vehicles is: input the environmental information collected by unmanned vehicles in real time, its own driving status, etc. The pre-trained model controls the unmanned vehicle to avoid obstacles according to the output of the model. [0003] In the prior art, the method of reinforcement learning is usually used for model training, and the model is obtained through continuous "trial and error" training. Specifically, when training the reinforcement learning model, the unmanned vehicle determines the reward according to the impac...

AI Technical Summary

Problems solved by technology

[0004] However, in the existing process of training the reinforcement learning model, it is usually determined that the model output is the correct control when the unmanned vehicle reaches the destination, and positive feedback is given. When a dangerous situation occurs during driving, it is determined that the model output is the wrong control. Give negative feedback, so usually only the rewards that reach the destination or the feedback of dangerous situations are effective rewards, that is, the rewards that make the model parameters converge, and most of the rewards are difficult to make the model converge during the driving process. The effective rewards obtained during the training process are relatively sparse, making model training costly and time-consuming

Images