Research On Muti-task Parallel Real-time Decision Model Of Auto Driving System

Although the decision-making technology used in the current auto driving system can enable smart vehicles to complete most of the independent decision-making tasks,with the substantial increase in the number of vehicles and tasks,the complexity of actual traffic environment is also increasing.In the face of parallel execution of multiple tasks,the system is still difficult to respond to the current real-time decision-making needs of vehicles,especially in the real-time performance guarantee of vehicle path planning,assisted driving,and emergency avoidance of dangers.In order to greatly improve the decision-making performance of intelligent vehicles in the multi-task parallel environment,it is necessary to design a real-time decision-making method for large-scale transportation systems,and further consider the decision conflicts caused by mutual interference between different types of tasks in the system.This thesis combines the high-efficiency computing power of auto driving system which focused on the study of multi-task dynamic decision-making algorithms to realize the real-time decision-making problem of the auto driving system.This thesis first describes the basic concepts and system architecture of the auto driving system,and expands to explain the shortcomings of multi-task parallel real-time decision-making in the current auto driving system,and further proposes the following solutions based on the nested neural network and the multi-granularity decision-making theory:1)By constructing the dynamic nested neural network,the structure of the learning model is adjusted online according to the attributes of the decision-making task to meet the training requirements of task dynamic allocation.Then combined with the Markov decision process,a multi-task dynamic offloading(MTDO)algorithm is designed and deployed in the nested neural network,which can adaptively adjust the multi-task parallel offloading strategy in real time to obtain high-quality decision results.Finally,this thesis conducted a simulation experiment to evaluate the proposed algorithm.The results show that the MTDO algorithm has a good performance in real-time decision-making in the auto driving system.2)Based on the theory of multi-granular decision-making,a multi-task parallel multi-granular collaborative decision(MPMCD)model is designed.Based on the theory of multi-granular decision-making,a multi-task parallel multi-granular collaborative decision(MPMCD)model is designed.This model uses a multi-granular information structure to improve the knowledge discovery ability of the decision-making process,and combines the sequential decision theory to build a neural network for multi-task parallel,which makes the decision-making in the automatic driving system further adapt to the fuzzy information source.Then a sequential decision based multi-task parallel real-time decision(SDMPRD)algorithm is designed by combining with deep Q-learning,which supports vehicle-road-cloud collaborative real-time decision-making with multi-task parallel.The simulation experiment compares with the MTDO algorithm and evaluates the decision performance standard of the SDMPRD algorithm.The results show that the SDMPRD algorithm is superior to the MTDO algorithm in terms of decision response time,accuracy and cost in a fuzzy environment.