PUCT

A tree search that starts from a hunch about which columns are worth trying, and is allowed to overrule the existing policy only when it is very sure.

The idea

The plain Monte Carlo tree search next door starts every position knowing nothing. Its first simulations are spent equally on all seven columns, including columns that any reasonable player would reject immediately, and it only learns to avoid them by trying them. In a game where a decision budget buys a few hundred simulations, that is a lot of the budget spent confirming the obvious.

PUCT — "predictor + upper confidence bounds applied to trees", the selection rule made famous by AlphaGo — adds a prior: a cheap opinion, formed before any simulation, about how plausible each column is. Columns the prior likes are tried first and often; columns it dislikes are not forgotten, but must earn attention. The prior here is not a neural network. It is three quick sampled probes per column using the repository's shared phase policy, turned into a distribution over columns by a softmax with a fixed temperature.

The second idea in this lab is caution. The search is not the player. The player is the existing phase policy, and the tree is only allowed to change its mind when the tree is emphatic — the challenger column must have been visited at least a minimum number of times, must hold a large share of all visits, and must lead by more than a fixed margin in value. Four such rules of increasing strictness are laid out side by side so a run would show how the outcome depends on how brave the override is.

How it works, step by step

1. Read the public position: board, visible next disc, moves until the next rise, terminal flag. Ask the shared phase policy for its action; that is the baseline.
2. Form a prior over the legal columns from three sampled probes each, flattened through a softmax with temperature 0.75.
3. Run 256 simulations over an eight-move horizon, with at most 4,096 stored positions. Selection scores each column by its average result plus an exploration term proportional to its prior and to how little it has been visited relative to its parent.
4. Compare the tree's favourite with the baseline. Accept the switch only if the challenger has at least 32 visits, more than 65% (or 80%) of the visit share, and a normalized advantage over the baseline above 0.10 (or 0.25). Otherwise play the baseline's column.

Simulation, gray-disc reveal and next-disc randomness each draw from a separate domain of the solver's own random stream, so the search's imagined futures cannot line up with the game's real ones.

What happened, in plain English

Nothing happened. That is the honest state of this approach: a finished experiment with no result.

It is not an oversight so much as a queue. The research roadmap places PUCT last in a sequence — first get a model that ranks moves safely, then let it override the reference conservatively, then batch and widen — and states the rule plainly: "Do not begin with learned action pruning. Every legal root action remains eligible until the model has independently demonstrated safe sibling ranking." A search that concentrates its budget by prior is exactly the kind of thing that rule is holding back until the ranking problem is solved.

Meanwhile the family's other results argue both for and against running it. Against: the scaled observable search showed that a bigger, better-directed tree became a better imitator of short-horizon values while getting worse at predicting long outcomes, and a prior directs the tree, it does not change what the playouts believe. For: this is a complete program with frozen gates whose cost an exploratory audit estimated at roughly one to three CPU-hours, and running an experiment that already exists is cheaper per unit of knowledge than writing a new one.

The technical record

There is no result record, no ledger entry and no artifact. What exists:

• Preregistered lanes in the source: an eight-game screen and an eight-game confirmation, on separate training seed ranges.
• One frozen search configuration — 256 simulations, horizon 8, exploration constant 1.25, at most 4,096 nodes — and four override gate profiles combining a 0.65 or 0.80 visit share with a 0.10 or 0.25 value margin at a minimum of 32 challenger visits.
• Separate random domains for simulations, reveals and next discs; a reflection-canonical treatment of mirrored positions; the shared public phase policy as both prior source and fallback.
• A build line in the reproducibility guide.

Audit 04 — exploratory, read-only — groups this lab with two others as "finished experiments with zero recorded outcome", estimates roughly one to three CPU-hours each, and adds a warning that matters more than the estimate: confirm with the coordinator that the preregistered seed ranges are still unconsumed before opening them. Eight games is a screen, not a measurement; even a clean pass here would be a reason to run a larger cohort, not a result about strength.

What this taught us, and what is still open

• "Not tried" is not "does not work." This page exists to keep those two apart. Nothing in this repository licenses the sentence "PUCT was tried on Drop7 and failed."
• The open question is narrow and cheap: does a prior plus a strict override gate beat the shared phase policy on its own preregistered eight-game screen? A run would produce the family's first result that is about deployment rather than about ranking stored positions.
• The prior is the weak point to watch. It is built from the same phase policy that the search is trying to improve, so a run that shows no switches would be telling you about the prior, not about PUCT.