trends.barnas.net

method · first seen 9 Oct 2026

imitation learning for branch-and-bound

Also: DAgger, BC/DAgger, DAgger distillation, dataset aggregation, behavioral cloning with DAgger

Learning solver branching and node selection policies from expert demonstrations with iterative data aggregation.

The label merges two older ideas. One is DAgger (dataset aggregation), an imitation learning method from about 2011 that repeatedly collects expert corrections on states the learner itself visits. The other is learning branching and node-selection policies for mixed-integer solvers by imitating strong branching, an approach studied for years. Only one source, a paper on sequence-structured branch-and-bound for MILP, is actually about solvers. The rest are robotics papers on humanoid control, VLA correction, mixture-of-experts gating for selective DAgger corrections and humanoid motion, plus imitation work on plasma control and MPC. These use behavior cloning or DAgger as a standard tool and do not form a new trend.

DAgger-style correction loops remain a common way to fix compounding errors in learned robot and control policies, so the term will keep appearing steadily. That reflects ongoing routine use rather than a new idea gaining ground.

Analyst note (established), 17 Nov 2026.

Mentions per week

0519 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 89 on 14 Nov 2026 · 8 mentions in 28 days · 8 independent groups · papers · first seen 5 weeks ago · acceleration +2.30

Earliest sources

Latest sources

Mentioned together with

vision-language-action modelmotion matchinghumanoid robot tennisbranch-and-boundmixed-integer linear programmingvision-language navigationhumanoid loco-manipulationwhole-body mobile manipulationbimanual dexterous manipulationmixture of experts