Board-static labels#
Status: adopted
The label is the hand eval of the collected board itself: no teacher replay at label time, no lookahead. This became the base recipe for every arm that followed.
Shared vocabulary in Glossary. This page adds:
| Term | Meaning |
|---|---|
| Board-static label | The hand eval of the collected board itself. |
| Static clone | A net trained on board-static labels, tracking the hand eval's ordering almost perfectly. |
The hypothesis#
If the goal is a fast, quantized stand-in for the hand eval inside the field term, the most faithful target is the hand eval value of the very board being evaluated.
What happened#
The clone, v7s, played the most hand-like of all arms and won 25% against production on the pools of its day. It also upstacked the most stubbornly, which is what exposed the real mechanism: see Field collapse.
The verdict#
Adopted as the base recipe. Its weakness is not the label but the distribution: train it on the states the system actually plays and the collapse disappears, which is the self-state pools fix.
Comments
No comments yet.
Sign in to comment.