Piloted driving vehicle training method based on virtual environment and depth double-Q network

Abstract

The invention relates to a piloted driving vehicle training method based on a virtual environment and a depth double-Q network. The method comprises: constructing a vehicle racing track virtual environment based on Unity; establishing a communication connection between the Unity and a piloted driving vehicle model under a Websocket protocol; establishing a Markov model that represents a vehicle piloted driving process, transforming a depth double-Q network algorithm to train a piloted driving model; configuring an entity vehicle soft hardware environment, migrating a trained vehicle piloted driving model; and testing an entity vehicle in a real racing track to train a piloted driving operation of the model. According to the method, a virtual environment is used for training a model, to implement robustness of a piloted driving training algorithm model on a complex environment and road condition. According to the method, by means of the depth double-Q network algorithm, a problem of high train complexity caused by a large assessment value of a Q value is avoided, so that a training process is simple and fast. The method has advantages of strong robustness, a fast speed, low costs and the like. The method can be applied to training and learning of unpiloted operation in fields such as intelligent traffic, aerospace, robots and the like.

Description

technical field [0001] The invention relates to a self-driving car training method based on a virtual environment and a deep double-Q network, belongs to the technical field of automation control, and specifically relates to a deep reinforcement learning algorithm. Background technique [0002] In recent years, with the rapid development of artificial intelligence technology, especially deep learning technology, major breakthroughs have been made in the fields of target recognition and intelligent control. Deep learning technology, such as convolutional neural network (CNN), is widely used in various types of vision-related problems, including the field of automatic driving, and the use of deep learning technology to achieve automatic driving is the mainstream of current research on automatic driving technology. At the same time, cars have entered thousands of households. With the increase of car ownership, the incidence of traffic safety accidents is also gradually increasi...

AI Technical Summary

Problems solved by technology

The inventive method needs artificial collection of training data according to the current feature model and environmental data, which requires a lot of artificial data collection work

The car trained by manually collected data has poor robustness, and the performance of the model will deteriorate after a slight change in the environment

Images