Addressing Policy Objectives of Traffic Control Using Evolutionary Algorithms
J Sha'Aban and B Heydecker, Centre for Transport Studies, UCL;
This paper describes the application of novel technologies of machine learning to develop a distributed responsive traffic signal control system. It uses evolutionary computing techniques for multi-agent systems adaptively to develop rules for on-line calculation of signal timings. This machine learning technique is based on a Learning Classifier System (LCS) that develops effective rules to deal with complex sequences of decisions and applies these rules to a traffic signal control system. The performance of this approach will be evaluated though comparison with established traffic control systems using the latest version of the microscopic simulation model SIGSIM. SIGSIM models the movement of individual vehicles in a signal-controlled road network. SIGSIM has an event-based capability where each event is processed in accordance with prevailing conditions. SIGSIM was designed for the testing and evaluation of signal control strategies and includes fixed time and System D vehicle actuated; it can also accommodate novel ones. There are a number of detector technologies available in SIGSIM, including inductive loops and microwaves. Learning Classifier Systems are a form of evolutionary algorithms that evolve simple condition/action rules, over a learning period during which reward values are fed back to the LCS from the simulation in respect of the policy
The aim of this is to generate adaptive control policies to facilitate a range of suitable behaviours. This approach to performance enhancement has the advantage that it is not specific to any particular objective or form of primary data. The integration of the LCS into SIGSIM will ultimately lead to the development of a fully distributed traffic responsive control system.
The integrated simulator implements responsive traffic signal control. SIGSIM provides simulated detector data and measures of performance. The LCS uses these measures to develop signal control strategies and gauge the efficacy of the generated signal timings according to how well the system performs. Initial investigations have used a measure of queue lengths at each junction as a local evaluation metric that corresponds closely to the mean rate of delay. A number of alternative evaluation metrics will be investigated for suitability when the traffic responsive system is developed further. Initial testing has used small sized networks and low complexity traffic flow patterns. Over the programme of development and evaluation these tests will progress to include larger size networks and increasing complexity traffic flow patterns. These results will show how well the LCS performs against fixed time and System D vehicle actuated, and will be presented in the paper.
Future investigations with the responsive traffic signal control LCS will include a series of tests using a range of test networks to investigate the different kinds of control strategies. These will include:
* using time-varying demand profiles to represent busy periods that cause transient overload,
* including temporary reductions in capacity due to incidents.
These tests will be undertaken using a range of different objectives and kinds of detector data to investigate the sensitivity of the approach to this and the good things about them.
Association for European Transport