Optimal Driving Policies Emergency For Large-scale Autonomous Vehicles Based On Multi-objective Co-evolutionary Algorithms

With the rapid development of automotive driving and connected cars,vehicles gradually obtain the capabilities of communication,perception and decision-making.Besides,vehicles and road infrastructures can interact directly or indirectly to each other.The transportation system presents the characteristics of large scale,high dynamics,high variability and uncertainty.Traditional traffic systems are difficult to respond to traffic conditions in a timely manner and make reasonable planning or scheduling for vehicles.With a centralized optimized scheduling method,the computational complexity will increase exponentially with the number of intersections and vehicles,complexity of road networks and possible conflicts between vehicles.Even if the computing power in the future is large enough,it is difficult for the centralized servers to respond to requests in real time due to the delay,loss,and transmission time.As the highly dynamic nature of the transportation network and the instantaneous changes in vehicle response,it is almost impossible to process real-world traffic events and dispatch vehicles in the planning network in a timely and centralized manner.Because traditional evolutionary algorithms(genetic algorithms,etc.)and some intelligent optimization algorithms(such as ant colony algorithm,particle swarm optimization algorithm,simulated annealing algorithm,etc.)contain only one population,the impact of competitive cooperation cannot be fully considered,and it is not easy to converge when the scale is expanded or the convergence speed is slow.On the other hand,the computational complexity of centralized optimization algorithms is too high to handle large-scale traffic optimization problems.Therefore,based on the methods of cooperation and competition,co-evolution algorithms are expected to be used to perform adaptive evolution processes to deal with various large-scale traffic scenarios.This thesis proposes an algorithm that combines co-evolution and swarm intelligence.Based on the driving strategy space,we construct multiple cooperative and competing populations,design a multi-objective reward function,make co-evolution process in a simulated environment and observe emergence of the dominant driving strategies.We have established a grid road model and a vehicle kinematics model in which each vehicle interacts in an indirect interaction manner,making the computational complexity linear with the simulated vehicle size.On this basis,we introduce the moderate strategies based on the q-state model and describe the moderate strategies with a lane change factor(the probability of lane change when there is not enough space),which enriches the autonomous driving strategy environment.In addition,we also analyze the stability and profit of strategic alliances based on cooperative game theory and give the lower band of strategy performance under various traffic density conditions.In the simulated environment,experiments on different traffic flow densities(the number of vehicles is 170 and 340 respectively)were carried out on a 4.5-km road.The results show that under different traffic flow densities,the winning strategy is different.Rational strategies can achieve the best results under higher traffic flow density.Aggressive strategy cause the highest accidents with a slightly higher average speed.The conservative driving strategy performed the worst and was quickly eliminated.At low traffic flow density,conservative strategy are quickly eliminated due to their low speed characteristics;radical strategies cause very few accidents and the average speed is at a high level;Rational strategies hardly cause traffic accidents and the speed is similar to aggressive strategy,so both rational strategies and radical strategies perform well in the evolution process.After adopting the best strategy,the total number of accidents and the average speed have been greatly improved.The number of accidents tends to 0 and the accident rate has been reduced to below 0.005%.The average speed has increased to 16.3m/s,improved by 30%.After introducing the moderate strategies based on the q-state model,it was found that the overall reward function score was further improved,the accident rate and average speed were further improved.This result has good enlightenment for designing driving strategies used future intelligent transportation systems and intelligent vehicles.Since the computational complexity of our method is linearly related to the number of vehicles and can be processed in parallel,the method can be extended to explore the optimal driving strategy for urban traffic that contains millions of autonomous vehicles.