Title: DEVELOPMENT AND EVALUATION OF A MULTI-AGENT APPROACH TO TRAFFIC SIGNAL CONTROL USING TRAFFIC SIMULATION
Abstract: This thesis is aimed at demonstrating the feasibility of using recent software development techniques based on distributed agent technologies to develop a decentralised multi-agent traffic signal control system. The motivation for this work stemmed from a recognition of the limitations of existing traffic signal control methodologies and an acknowledgement that traffic control is a large-scale, complex, dynamic problem with highly nonlinear characteristics. A review of existing systems revealed a number of challenges and limitations and showed that the majority operated principally on pre-defined plans which were suitable only for pre-determined situations. As a result, they failed to control unexpected conditions when traffic flows changed rapidly. The performance of existing systems was also found to suffer during over-saturated conditions and the main challenge in this thesis was formulated as the need to shift from controlling near-saturated conditions to the requirement for traffic control during over-saturated conditions. This thesis adopted a traffic simulation approach for the evaluation of the proposed multi-agent traffic control system. It also introduced a number of contributions in the calibration and validation process using techniques for queue discharge headways collected from a field survey in addition to comparing the intersection performance measures obtained from the traffic simulator with those obtained from the well-known traffic signal optimisation tools such as the aaSIDRA software. The calibrated parameters were also evaluated and verified using macroscopic traffic flow theory which showed that the flow-density, speed-density and speed-flow relationships can be adequately reproduced using the calibrated and chosen parameters. The calibration and validation results showed overall model errors in the range of 10 to 14 percent, which was a very good result that provided a high level of confidence in the ability of the model to reproduce field conditions. In the multi-agent traffic signal control system, each traffic signal controller was modelled as an agent software and was equipped with two optimisation models: a signal timing optimisation model called “Cycle Optimiser” and a traffic signal coordination algorithm called “Offset Optimiser”. The agents were modelled as distributed systems with ability to collaborate and coordinate activities to achieve local and global goals. Each agent (intersection) was comprised of knowledge about intersection parameters, and had the abilities to generate the optimal control plans and collaborate with other agents through a traffic signal optimisation model. Each agent was also assigned a set of individual preferences or settings including objectives, a set of pre-determined plans and algorithms to generate plans. Each control agent received traffic flow data at the end of every cycle from vehicle loop detectors at the intersection’s stop line and also received vehicle occupancy data from the same loops at the beginning of the cycle. Traffic flow data was then used by the signal optimisation model to generate the control plans based on real-time traffic conditions. The multi-agent control strategy was comparatively evaluated with a fixed time control system under two traffic conditions: recurring and non-recurring congestion. It was shown to be superior to fixed time control systems for isolated intersections. It increased the intersection flows by 11.64 percent and increased the travelling speed by 23.23 percent. In addition, it also reduced the number of stops by 43.36 percent and consequently reduced the intersection delay by 44.15 percent. The agent control system was also superior to fixed time control systems for coordinated intersections. It resulted in increasing the network flows by 5.05 percent and average travel speeds by 5.57 percent compared with a coordinated fixed time system. During incident conditions, it was also superior to fixed time control systems. Its performance was further enhanced by providing mid-block detectors. For minor incident cases, these benefits included increases of 3.64 and 4.14 percent in traffic flows and average speeds, respectively. These were accompanied by decreases of 10.50 and 11.70 percent in average delays and queue lengths, respectively. A comparative evaluation between the agent control system and SCATS-generated signal timing plans for a case study in Bandung, Indonesia, showed that for the AM peak, an unsupervised agent-based controller increased the overall network flows by 2.85 percent and the average travel speed by 3.23 percent. It also reduced the number of stops by 2.21 percent and the network delays by 4.84 percent, compared with the SCATS plans. Overall, the multi-agent control system was shown to reduce the average delay on the network by at least 25.74 percent when compared with the SCATS plans. The benefits were larger at 28.33 percent when the SCATS plans were compared with a coordinated multi-agent traffic signal control system. These combined findings clearly demonstrate the potential of agent-based traffic signal control systems in enhancing performance, reducing congestion and improving environmental air quality. This thesis has successfully achieved its stated objectives by demonstrating the feasibility of applying distributed systems of multi-agent algorithms to traffic signal control and by evaluating its performance and comparing it to optimised fixed-time control systems under recurring and non-recurring (incident) traffic conditions using microscopic traffic simulation tools. The study also successfully achieved a number of secondary objectives including advancing the state of knowledge in the calibration and validation of traffic simulation models; enhancing the procedures for signal coordination; improving the methods for reliable demand prediction and forecasting; and formulating general methodologies and frameworks that can be used for evaluating the performance of Intelligent Transport Systems. Finally, the thesis identified a number of issues as potential research areas in this field. These included the potential to improve the accuracy of the algorithm using additional real-time data such as vehicle occupancy from loop detectors; the need to use artificial intelligence approaches based on neural networks for speed estimation and short-term traffic forecasting; interfacing the agent system to an automated incident detection algorithm so that non-recurring congestion is more easily identified, and broadening the coverage of the road networks, testing a much large number of intersections and testing the computational efficiency of the multi-agent traffic signal control systems.
Publication Year: 2007
Publication Date: 2007-01-01
Language: en
Type: dissertation
Indexed In: ['crossref']
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot