Bayesian Optimization on Causal Models for Railway delay management

Abstract

This project aimed to solve a problem for ProRail, what action should a traffic controller take to minimize delays in an active railway system? and an AI system was developed to do just that. Using a structured causal model to estimate future delays based on predicted train interactions, a separate Bayesian optimization algorithm would read and intervene on the SCM to find the actions that produced the lowest delays. This project compared two different acquisition functions: log expected improvement and lower confidence bound, each with their own appropriate termination criteria and found that although there was no difference in accuracy between the two, the log EI algorithm proved to have a better performance on the tests run. There was also a significant improvement in using the algorithm than not doing any intervention at all in a system where delays exist. Finally the choice of intervention seems to match the rules ProRail currently has for action the traffic controllers must make.