Research On Autonomous Decision-making Algorithm For Autonomous Vehicles Based On Meta Reinforcement Learning

With the rapid development of artificial intelligence,unmanned driving has become a research hotspot in the automotive industry.Safe and reliable intelligent driving strategies can free the driver’s hands and improve the driving experience.It has extremely important practical significance.Nowadays,mainstream unmanned vehicle decision-making algorithms are mostly model-based methods.Although the algorithm is highly interpretable,it does not have the ability to adapt to learning.The decision-making effect is limited by the model’s ability to express,and the algorithm is not robust in multiple scenarios.In response to the above problems,this article has done the following work to improve the autonomous driving behavior decision system.First,use deep reinforcement learning to replace traditional decision-making algorithms.In this paper,model-free,heterogeneous strategy and adaptive deep deterministic strategy algorithm(DDPG)are used to realize the horizontal and vertical decision control of driverless driving.DDPG can accept high-dimensional perception data input and realize continuous decision output.Unmanned vehicle decision-making system based on DDPG algorithm can self-learn by interacting with virtual road environment,and it has good robustness in different scenarios.Second,when deep reinforcement learning is applied to unmanned vehicle decisionmaking,the training of unmanned vehicles needs to go through a blind trial and error stage,resulting in low sample efficiency and slow model training.In view of the above problems,this paper introduces the concept of meta-learning and proposes meta-Deterministic strategy algorithm Meta-DDPG.The algorithm can generate a set of excellent initialization parameters that can be used for any similar tasks,so that the model has the initial decision-making ability,and on this basis,specific scenarios or tasks can be targeted for training.The Meta-DDPG algorithm proposed in this paper can effectively improve the model convergence speed and improve the model robustness.The research focus of this paper is: in view of the shortcomings of blind trial and error based on DDPG-based automated driving strategies,a meta-depth deterministic strategy gradient model is designed.In the early stage of deep reinforcement learning training,an integrated strategy of meta-learning was introduced to improve the training speed of the model.Considering the important influence of the reward function on the model decision-making effect,the reward function is designed based on the safety and smoothness requirements of unmanned driving,combined with the generalization requirements of multiple scenarios.