I have recently been doing a literature survey of research into train timetabling and re-scheduling for a potential contract (using machine learning et al). Dealing with train delays in real time is not an easy problem to solve; it's pretty much a NP complete problem to find the most optimal solution. So 'the best so far' is the goal given time constraints. So I'm not surprised at all about how one issue went and cascaded. Managing real world train networks is pretty complicated because of the interconnection of so many things - the train schedules, the crew schedules, managing the rolling stock, the passengers... and the profit/loss of the train company.