New AI Trick Solves Tough Problems With Far Fewer Tries
Could mean cheaper drugs and faster car designs — using less computing power.
Researchers introduce a method called Iterative Sequential Transfer to tackle few-shot multiobjective multitask optimization. Instead of transferring knowledge across all tasks at once, it models multitask optimization as a sequence of sequential transfer optimization problems, concentrating evaluations on a single target per iteration. A likelihood-informed task prioritization mechanism aims to maximize transfer utility by identifying the task most likely ready for knowledge integration. The paper says this circumvents a bottleneck: knowledge transfer in multitask optimization relies on aligning elite solution distributions across tasks, which is hard when restricted evaluation budgets impede identifying those distributions — a challenge made worse in multiobjective multitask problems, where each optimizer must approximate a continuous Pareto manifold rather than a single optimal point. The authors report that empirical results on benchmark and real-world problems verify the method's effectiveness under tight budgets.
- Instead of solving many hard problems at once, the new method solves them one at a time — in the smartest possible order.
- It targets "few-shot" situations, where you only get a handful of expensive attempts to find a good answer.
- Tested on standard academic problems and a few real-world ones, it beat existing methods when budgets were tight.
Why It Matters
Cheaper, faster engineering and drug design — fewer costly tests needed to find good solutions.