Line labels#

Status: refuted

Two arms labeled each board with a statistic along the teacher's full line: the mean, v5, and the minimum, v6, of the hand eval values seen along it.

Shared vocabulary in Glossary. This page adds:

The hypothesis#

A line statistic sees several placements ahead, so the net should learn consequences the single-board value cannot express, while staying bounded: a mean or minimum over boards cannot grow with horizon the way accumulated reward does.

What happened#

The minimum anticorrelated with the hand eval: boards along the teacher's line are often temporary sacrifices, so the minimum rewards lines that dive. The mean was safer but still lost heavily in duels.

The verdict#

Refuted. Statistics along a line still re-import horizon-dependent structure in a softer form. The terminal value of the line is a different quantity, bounded and without the dive incentive; Blending tested it and refuted it too.