Distributed Control of Collaborating and Learning Expert Systems for Street Traffic Signals

Abstract

The paper describes our long-term activity aimed at the control of traffic signals by a network of distributed processors situated at street intersections. Every processor runs an identical expert system (possibly with a somewhat different knowledge base) and communicates directly with the four adjacent processors. Messages can reach also indefinitely distant processors, modulated by the needs of intervening ones. The information transmitted can be raw data, processed information or expert advice. The rule-base of the expert systems has a natural segmentation, corresponding to different prevailing traffic patterns and the respective control strategies. Multi-dimensional learning programs optimize both the hierarchy of the rules and the parameters embedded in individual rules. Different measures of effectiveness can be selected as the criterion for optimization. Traffic scenarios are automatically generated for a ‘characteristic period’ — for a certain part of the day (e.g., morning rush hours), a certain type of day in the week (e.g., regular work day), a certain season of the year (e.g., vacation time). The results of our first implementation, a running prototype simulation program, has proven the feasibility and utility of the approach.