Research On Autonomous Decision Making For Intelligent Vehicles Based On Reinforcement Learning

The development of intelligent driving vehicles technology is of great significance to ensure the safety of vehicles and relieve the traffic pressure of the city.Intelligent driving vehicles is an important application area of artificial intelligence.The independent decision-making ability of intelligent vehicles can represent the level of its intelligent level.How to use the machine learning method to improve the independent decision-making ability of intelligent vehicles and improve the safety of intelligent driving has become to be a focal and difficult spot in the field of intelligent driving research.Based on current research about intelligent vehicles driving technology,lane change decision for intelligent vehicles in a structured road environment has been researched deeply in the paper.Currently,the current rule-based decision-making method is not comprehensive on the rule base and the method based on statistical decision-making demand a lot of data.So as to have a difficult to adapt the dynamic structured road.For the lane change decision a structured road environment,a method based on reinforcement learning has been proposed for intelligent vehicles.Intelligent vehicles can be more intelligent by the method.The main contributions of this paper are as follows:(1)A algorithm based on Multi-Kernel Least Squares Policy Iteration(MKLSPI)is proposed for reinforcement learning with large scale and continuous state space.The algorithm uses the linear weighting of multiple kernel functions to construct the feature and is used to approximate the value function in the process of the policy iteration.More parameters often need to set and adjusted artificially when using the reinforcement learning algorithm to solve large-scale or continuous state space problems.In this paper,multi-kernel functions are used to construct features automatically and fit the optimal features by linear weighting.The advantage of the method is to reduce the number of parameters that need to be set manually to improve the flexibility of the algorithm.The proposed algothrim is simulated by two classical learning control problems of Mountain-car and inverted pendulum.The simulation results show that the algorithm has good abilities of feature representation and generalization,which can effectively reduce the difficulty of optimizing the learning parameters while ensuring the performance of the algorithm.(2)A driving independent decision-making method based on MKLSPI is proposed for intelligent vehicles.In this method,the decision-making problem is modeled as a Markov Decision Process(MDP)with continuous state space.Then,the MKLSPI is used to train and learn the optimal or near-optimal decision-making policy.The method is based on data-driven and obtain the driving experience from the sample data.So,it can improve the ability of adaptive the environment and learning for intelligent vehicles.A driving decision making simulation platform based on the autonomous vehicles HQ3,and on which the proposed method is evaluated.(3)In order to fit in with the current highway environment,a intelligent vehicles driving decision-making simulation system based on multi-lane highway environment is designed and achieved in this paper.The problem of intelligent vehicles driving decision-making is solved by the method based on multi-kernel API proposed in this paper while the dimension of the MDP increases.The method is tested in the highway simulation environment and the results show that the intelligent vehicle using the decision policy learned from samples data can run safely in the dynamic traffic environment.It is meaningful to solve the problem of intelligent vehicles driving decision-making by reinforcement learning.