conf-schedule-optimization
Conference Schedule Optimization
Choosing what to attend is a constrained optimization problem, and treating it as one is what separates a real schedule from a wishlist. Events are variables; time slots are their domain; hard constraints (you cannot be in two rooms at once, you need minutes to walk between them, a capped workshop may be full, you said you must catch the opening keynote) define what is feasible; and a weighted objective over the feasible plans defines what is good. The output is the best feasible plan plus an honest account of what it could not decide and what it gave up to score well.
The objective is genuinely multi-term and the terms are in real conflict: maximizing interest fights breadth (seeing across themes), both fight pacing (a day of six back-to-back talks with fifteen-minute gaps is worse than five with a real break), and all three fight serendipity (deliberately leaving room for the unplanned good thing). You do not resolve this by pretending one term dominates. You jointly optimize a weighted sum whose weights are the attendee's — handed over from the preference profile, never invented here.
Two disciplines keep it honest. Conflict-surfacing: when two talks overlap and both score high and the profile cannot break the tie, the system flags it for the person rather than quietly choosing — that flag is the natural output of a model that tracks its own uncertainty. And the Goodhart caution: any score you optimize produces a substitution effect somewhere the score did not name (optimize raw interest and you quietly sacrifice breaks; optimize breadth and you skip the one talk they cared about most). So the weights stay user-owned, the conflicts get surfaced, and the schedule reports what it traded away.