MARCS Multi-agent Railway Control System

Abstract

Previous research works have demonstrated that traffic control models based on the comparison between an historical archive of information and current traffic conditions tend to produce better results, usually by improving the system’s proactivity behavior. Based on this assumption, we present in this paper MARCS – Multi-Agent Railway Control System, a multi-agent system for communications based trains traffic control. For this purpose we have developed a system infrastructure based on an architecture composed of two independent layers: “Control” and “Learning”.

“Control” layer is responsible for traffic supervision, regulation, security and fluidity, including three distinct agent types: “Supervisor”, “Train” and “Station”.

The “Learning” layer, using situations accumulated by the “Control” layer, will infer rules that can improve traffic control processes, minimizing waiting time and stop orders sent for each train. At this moment, inferred rules seem like:

“At T₁ moment, when a train is located at P ₁ = (x1, y1) with destination E ₁ and another one is at P2 = (x2, y2) with destination E ₂, a traffic conflict in L ₁ after t ₁ seconds” will occur.

Rules of this kind are transmitted to the control system to be taken into account whenever a similar traffic situation is to occur. In the learning process we apply an unsupervised learning algorithm (APRIORI).