A new AI system called Delphos can specify and test transport choice models nearly as well as expert modellers, according to a paper posted this week.
Researchers built Delphos as a multitask reinforcement learning framework that treats model specification as a sequence of decisions, applying modelling actions and getting feedback from an estimation environment. It represents utility specifications as sets of modelling terms using a DeepSet-Q architecture, letting a single policy transfer what it learns across datasets that have different variables. The team trained Delphos on nine transport choice datasets, where it consistently outperformed single-task agents trained from scratch on one dataset at a time. Tested without further training on two datasets it had never seen, Swissmetro and Decisions, the same agent produced competitive specifications in under 20 minutes on a standard CPU, beating a VNS metaheuristic on Swissmetro and matching a published expert-built specification on Decisions.
Choice model specification is a time-consuming task, since modellers have to juggle goodness-of-fit, parsimony, and behavioural plausibility across multiple candidate models. A tool that reuses experience across datasets and narrows the search space could cut a lot of that manual trial-and-error, all without taking the modeller out of the loop for diagnosis and final selection.
Matching a human-built specification and beating a classic metaheuristic on two unseen datasets is a genuinely useful result. But "competitive with experts" is not the same as "better than experts": for now, Delphos looks like a fast first draft, not a replacement modeller.