Multiple Unmanned Vehicles Intelligent Decision Based On Reinforcement Learning

In recent years,deep learning algorithms represented by convolutional neural network(CNN)have shown remarkable performance in many tasks such as target recognition and detection.Learning ability of complex tasks is very important to the development of AI.Generally speaking,it is difficult for neural to do this.It is generally believed that catastrophic forgetting is an inevitable feature of connectionism model.At present,the AI system based on deep neural network is designed for learning narrow-range tasks,so the trained agents do not have continuous learning.The key step to achieve general AI is to acquire the ability of continuous learning,that is to say,an agent must be able to learn how to perform new tasks without forgetting the execution methods of old tasks.In order to train sustainable learning agents,the urgent problem to be solved is to solve the catastrophic forgetting problem.Unmanned vehicles need to drive in various complex scenarios,and need to ensure the safety of vehicles in different scene.So we hope that unmanned vehicles can constantly acquire new skills while performing tasks in various scene,without forgetting the skills that have been trained and mastered.In this paper,a memory mechanism is added to solve the catastrophic forgetting problem in the test environment of multiple unmanned vehicles to prevent potential safety hazards caused by knowledge forgetting in different scenarios.Here we present a multi-vehicle scene learning framework,which allows unmanned vehicles to memorize different scenes.In order to prevent the unmanned vehicle from forgetting the knowledge learned in the original scene in the learning process of the new scene,so that the unmanned vehicle can learn the new task better,and at the same time less forget the knowledge learned in the previous taskFor multi-vehicle environment,we find that the driving strategies learned by different unmanned vehicles have commonality and repetition of learning process.Therefore,this paper proposes an accelerated training method based on experience sharing,which can improve the learning effect of unmanned vehicles from two aspects of time-consuming and convergence speed by sharing the experience of multiple unmanned vehicles in the initial learning process.The similarity of vehicle strategy is used to transfer the basic knowledge network to improve the learning starting point of unmanned vehicles in other environments.Finally,the simulation experiment of the proposed solution in multi-vehicle environment shows that:(1)The accelerated training method based on experience sharing proposed in this paper is superior to vehicle independent learning in convergence speed and time-consuming,and its learning effect is comparable to that of vehicle independent learning;(2)The scene learning model constructed in this paper can achieve learning effect and traditional deep reinforcement learning.At the same time,in multi-scene scenarios,unmanned vehicles have better security performance than the original deep reinforcement learning algorithm in scene conversion,and no forgetting occurs between different scenarios.