The reliability of railway traffic is commonly evaluated with train punctuality, where the
deviations of actual train arrivals/departures and train arrivals/departures published in the
timetable are compared. Minor train delays can be mitigated or even eliminated with running
time supplements, while major delays can lead to so-called secondary delays of other trains
on the network. Railway lines with high capacity utilization are more likely subject to delays,
since a greater number of trains means a larger number of potential conflicts and more
interactions between trains. Consequently, the secondary delays are harder to limit. Railway
manager and carrier personnel are responsible for safe, undisturbed and punctual railway
traffic. But unforeseen events can lead to delays, which calls for train rescheduling, where
new train arrivals and departures are calculated. Train rescheduling is a complex
optimization problem, currently solved based on dispatcher’s expert knowledge. With the
increasing number of trains the complexity of the problem grows, the need for a decision
support system increases. Train rescheduling is considered an NP-complete problem, where
conventional mathematical and computer optimization methods fail to find the optimal
solution, but artificial intelligence approaches have some measure of success. In this
dissertation an algorithm for train rescheduling based on reinforcement learning, more
precisely Q-learning, was developed. The Q-learning agent learns from rewards and
punishments received from the environment, and looks for the optimal train dispatching
strategy depending on the objective function.
|
