Glossary · Term

DAgger

← all terms

Definition

Plain language

A training fix where the learner drives, gets stuck, and an expert says what it should have done at exactly those stuck moments.

As stated in the literature

Dataset Aggregation, a 2011 imitation-learning algorithm that iteratively collects expert corrections at the states the learner actually visits, addressing the distribution-shift weakness of plain behavior cloning; impractical for GUI agents because it requires a cheap on-demand expert.

Why it matters: It fixes the blind spot of plain imitation by teaching recovery from the learner's own mistakes, though it needs an expert available on demand.

For example, a learner attempts a task, and whenever it wanders into a tricky spot, an expert steps in to say exactly what the right move there would have been.

Heard on the show

“… There's a classic result in imitation learning, a method called DAgger from back in twenty-eleven, and the idea is precisely that: let the learner drive, see where it gets …”
Episode 155 — Why a Flawless Demo Makes a Worse Computer-Using Agent, And the Fix

Mentioned in 1 episode

  1. 155
    Why a Flawless Demo Makes a Worse Computer-Using Agent, And the Fix

Related terms