Research On Reinforcement Learning And Its Applications In Urban Traffic Signal Control

In our modern society, traffic jams are ubiquitous and getting worse in and aroundurban areas. Traffic signal control is one of main means of controlling road traffic, soimproving and optimizing traffic signal control methods is a very effective way to solvetraffic jams. Due to the high dynamic and randomicity of traffic flow, furthermore,traffic signal control is fundamentally a problem of sequential decision making, soreinforcement learning is suitable for the control of traffic flow. The traffic signalcontrol system of the traffic network is viewed as a Multi-Agent System where eachtraffic signal controller controls one intersection, reinforcement learning can be appliedto this Multi-Agent System, so one of the main means to solve traffic congestions andtraffic jams is to design and develop reinforcement learning-based adaptive traffic signalcontrol methods.In this thesis, first, model-based reinforcement learning in traffic signal control isexplored and studied. Improve TC1to give birth to TCSG. In TCSG, the length of everyvehicle is considered to be calculated to the reasonable congestion of destination lanesof the vehicles which is used to implement the sharing traffic condition coordinationamong Multi-Agent i.e. traffic signal controllers. In GLD, a mass of simulationexperiments show that TCSG outperforms TC1.Secondly, model-free reinforcement learning in traffic signal control is explored andstudied. Design and develop new traffic signal control methods i.e. DMFQ, QSGWE,DMFQV, DMFS and SSGWE, they are vehicle modeling-based to slove extremely largestate space, they also achieve sharing coordination among traffic signal controllers, theyalso adopt―Waiting Voting Mechanism‖to predict and select the optimal action ofevery traffic signal controller–Agent. A number of experiments simulated in GLD showthat our methods outperform Pretimed Control and TC1.Finally, conclude the work of this thesis, point out the problems in the research of mytraffic signal control methods, bring forward the future research in traffic signal control.