Research Of Reinforcement Learning Algorithm For Unmanned Driving

Unmanned driving is a hot research field in the current world academic field,and there are still a lot of problems that have not been solved yet,which have important research significance.Aiming at the problem of the unmanned driving based on deep learning server platform,we finish a methed of unmanned driving for vehicle's driving.The main contents of this paper are as follows:Based on deep neural network,aiming at the problem of vehicle driving in virtual environment,we complete the design of reinforcement learning algorithm for unmanned driving.The reinforcement learning method is an important method to study under the unknown condition of the model,and many RL methods are regarded a goal as the longterm optimization goal.When the continuous space is too big and the state is unknown,the deep reinforcement learning method can be used to construct the deep network and realize the control decision of the target.The controling algorithm provides the target operation information for the control system.In the process of the target operation,the algorithm can provide real-time feedback on the operation of the current vehicle,and can conduct trial and error learning according to the situation.Comparing with the traditional deep learning,we use the deep reinforcement learning algorithm to learn strategy,which is closer to the way of human learning.When we train the deep neural network,we find the vehicle's driving is not smooth and the training time is too long,in this case we proposed a deep reinforcement learning algorithm which is based on residual learning.The biggest improvement is to add the residual network to the deep network.For enhancing the ability of network learning,the number of layers in the network can theoretically be increased indefinitely.However,as the number of network layers deepens,the learning performance of the network becomes worse.In order to avoid this situation,the residual learning is introduced into this algorithm.After outputting the deep network's value,the residual network is used to evaluate the value in advance,and then filter out the unreasonable value when the network output is changed.Then the boundary judgment was used and the value was reassigned.Adding residual study one can make goals in the process of the vehicle is more smooth,and the algorithm can effectively lower the training time and shorten the training cycle.The important work of this paper is to rely on the deep learning server platform in the laboratory and realize a reinforcement learning method that can provide decisionmaking for vehicles in the virtual environment.Using this deep learning server platform,the simulation experiment can be carried out in virtual environment.Due to the high risk,long-time cycle,high price and difficult to realize the special conditions,these problems can be ignored in the simulation environment.Therefore,it is necessary to simulate driving simulation experiment in virtual environment.Because the problem of control and decision-making in vehicle unmanned driving is a complex sequential decision problem with large-scale continuous space and multiple optimization goals,the optimization goal is difficult to solve with traditional dynamic programming method.because of this,we put forward the reinforcement learning algorithm for unmanned driving in this paper.