Successor-static labels#

Status: superseded

The label is the hand eval of the board after the teacher's chosen move: a one-placement lookahead. This produced the first deployable net, v4.

Shared vocabulary in Glossary. This page adds:

The hypothesis#

One placement of lookahead makes the label bounded, a single board value, while keeping some of the teacher's selection knowledge: the label reflects the board the teacher chose to reach, not just the board it started from.

What happened#

The recipe was bounded and trained cleanly. The resulting net won 18% of games against production, the first deployable distillation. Over time it became clear the recipe is a close clone of the hand eval's ordering, which caps what the net can add.

The verdict#

Superseded by the board-static label, which trains on the recorded board itself and needs no teacher replay at label time. The clone ceiling it revealed is the subject of Blending.