The vehicle lane change decision is a greet challenge for autonomous driving in the dynamic traffic environment. This problem is a sequential decision problem with high dimensional and continuous state spaces. Traditional methods such as the rule-based method is difficult to solve such complex optimization decision problems. Recently, solving the problem by using the reinforcement learning method has become a hot topic in the field of autonomous driving. In this paper, a method based on multi-kernels least squares policy iteration (MKLSPI) is proposed to solve the lane change decision problem. The weighted multi-kernels function with different widths can solve the problem that the width of single kernel is difficult to determine and this reduces the difficulty of adjustment and optimization of learning parameters. In this way, the final policy can be obtained by training and learning from the samples data. The efficiency of the policy is verified in the highway simulation environment.