Research On Adaptation Technology Between Virtual Driving Model And Environment

To improve the autonomous navigation ability of the virtual driving model is not only the key problem to improve its adaptability to the environment, but also to realize its application in complex and unknown environment. Because the reinforcement learning system can learn from environment, it has no need of prior knowledge and is a form of non-tutor learning method, it has been widely used in artificial intelligence field. And artificial neural network is a valid method to solve the generalization problem for the continuous state and action pairs in the reinforcement learning method.The reinforcement learning algorithm and artificial neural network are analyzed in detail. With Q-learning, one kind of reinforcement learning methods, to realize the autonomous navigation for the virtual driving model, and with BP neural network to solve the generalization problem for action selection. In this thesis, collision detection is one of the elements to design the state and the value function of the reinforcement learning, and it is also the key technology to improve the reality for the virtual driving model, the Sphere method is used to solve the collision detection between the virtual driving model and obstacles.The method of getting 3DS data by OpenGL is used to create the virtual driving model, so that the model is more living and the modeling time is reduced. Creating its dynamics model is an important problem for virtual driving, a rational dynamics model is founded for the virtual driving model according to automobile theory, and numerical method is used to analyze the automobile motion. In addition, the dynamical scheduling for the scene data is discussed, a visibility culling method named virtual occluder is adopted, schedule the scene data according the visibility of the virtual occluders, so that to realize the dynamical scheduling for the scene data.Visual C++ 6.0 is adopted as the development tool, combined reinforcement learning and artificial neural network to develop a virtual driving model that can adapt to the environment, the model is tested in the scene developed with OpenGL library in the prior term, and the experiment has accomplished the autonomous navigation for the virtual driving model, the result shows that the algorithm is available and feasible.