Vehicle convergence control method based on deep reinforcement learning algorithm

Abstract

The invention discloses a vehicle convergence control method based on a deep reinforcement learning algorithm, and the method comprises the steps: decomposing a vehicle convergence scene into a plurality of problem points, and dividing the problem points into two types: problem points suitable for deep reinforcement learning training and problem points not suitable for deep reinforcement learningtraining; for the divided problem points suitable for deep reinforcement learning training, extracting feature values in a vehicle convergence scene and inputting the feature values into a deep reinforcement learning algorithm for training to obtain predicted convergence information of each problem point; and for the divided problem points which are not suitable for deep reinforcement learning training, directly extracting feature values in the vehicle convergence scene by utilizing a feature extraction method, combining the feature values with the predicted convergence information of the obtained problem points to carry out convergence logic judgment, obtaining a convergence control result, and executing the convergence control result. According to the method, the training difficulty of deep reinforcement learning is simplified, the convergence logic judgment degree is deepened, the result of the convergence control process is more accurate, and the safety of vehicles in automatic convergence control is improved.

Description

technical field [0001] The invention relates to a vehicle merging control method based on a deep reinforcement learning algorithm, and belongs to the technical field of vehicle driving control. Background technique [0002] With the continuous breakthrough of artificial intelligence technology, after the combination of deep learning and machine learning in artificial intelligence technology in automatic driving, the level of automatic driving has also made significant progress. [0003] Deep reinforcement learning is an important direction of artificial intelligence. Its principle is that the agent interacts with the environment in the set environment, the agent makes actions in the environment, the environment rewards the actions, and the agent learns according to the actions. The principle It is equivalent to the human self-learning evolution process. When the algorithm is designed properly, it can theoretically drive better than humans. Therefore, there are infinite possi...

AI Technical Summary

Problems solved by technology

[0004] Reinforcement learning is an important branch of artificial intelligence, but the current level of development of deep reinforcement learning algorithms shows that it is suitable for processing tasks in simple scenarios, and the output of the algorithm is based on a probability model, which means that its output is uncertain

However, autonomous driving is faced with a complex scene, and autonomous driving has high safety requirements. The uncertainty of deep reinforcement learning has a great impact on its application in the field of autonomous driving.

[0005] Currently, if figure 1 As shown, the application of deep reinforcement learning in the field of automatic driving is mostly realized through the black box model of perception end-control end. There are big problems in the explainability and maintainability of the algorithm, which makes the entire automatic driving process a black box. There are many problems, which are not conducive to driving safety

[0006] For the lane intersection scene that is common in daily driving, if the deep reinforcement learning algorithm is directly used to train the automatic merging skills, the deep reinforcement learning algorithm needs to judge whether it is necessary to merge, when to merge, and the speed of the merge, etc. If there is a problem in the middle step , it will lead to the failure of the entire confluence, bringing a greater security risk

Images