---
title: Evolution
family: heuristic-search
summary: Let a search tune board-feature weights by playing thousands of games, keeping the best weights, and repeating.
status: rejected
evidence: ledger-recorded
reads: public
---

Stop arguing about how much each board feature is worth and let a search find
the numbers — play thousands of games with slightly different weights, keep
the ones that scored best, repeat.

<EvidenceLabel status="rejected" evidence="ledger-recorded" reads="public" />

## The intuition

Every hand evaluator in this repository contains a wall of constants: −900 for
peak height risk, +300 for quiet build options, −550 for adjacent 1s. Nobody
derived those numbers; somebody chose them. It is an obvious idea to let the
game itself choose them instead.

The method is derivative-free optimisation on complete games. Perturb the
weight vector, play whole games with each variant on the *same* seeds so that
luck cancels, rank the variants, move the distribution toward the winners, and
repeat. It needs no gradient, no labels, and no theory about why a feature
matters — only a lot of games.

Three programs here try three versions of that:

1. **Linear** — 40 features describing what a move did and what board it left,
   scored as a weighted sum.
2. **Nonlinear**: the same 40 features fed into a small 12-unit network with
   505 parameters, first taught to imitate a proper search and then evolved.
3. **Phase weights**. Instead of free parameters, seven interpretable groups
   of the existing reference leaf are scaled up or down together.

## How it works, step by step

1. **Extract features for each legal column**: what the move immediately did
   (score, clears, reveals, crack progress, chain depth, board clear, death,
   level advance) and what the resulting board looks like (occupancy and
   projected occupancy debt, heights and roughness, cover load and altitude,
   readiness and trench features, low-number clogs, outcome dispersion).
2. **Score each column** with the candidate weights and play the best.
3. **Play whole games** with every candidate on shared seeds, and rank
   candidates by a mixture of mean and median score — deliberately not by mean
   alone, so that one extraordinary chain cannot dominate a generation.
4. **Update the distribution** toward the best candidates and repeat for a
   fixed number of generations.
5. **Freeze the winner** and test it on games it was never fitted on.

Step 5 is where all three died.

## What happened

Each version learned its training games and gave the gain back on fresh ones.

The nonlinear version failed most informatively, because it was built as a
funnel that has to pass a checkpoint before it is allowed to spend real
compute. It first cloned a proper search's decisions, and the clone matched
almost all of the actions it was trained on while matching fewer than half of
the same teacher's actions on states it had not seen. That is a textbook
distribution shift, caught before a single evolutionary generation ran. It
then lost its four-game screen by half, and the funnel stopped itself.

The phase-weight version is the most interesting failure, because it was
*directionally right*. Six generations independently agreed that the reference
leaf should care more about cover debt and about trigger and rise readiness —
and a completely separate privileged audit later agreed with those same
directions. The frozen vector still lost its screen. Turning every weight in a
group up or down together cannot express the state- and action-conditional
interactions the game actually needs.

<TechnicalDetails title="The technical record">

The [experiment index](/docs/research/experiment-index) records all three as
**rejected, ledger-recorded**: *"each family failed its whole-game gate
despite inexpensive fitting or plausible feature directions."*

**Nonlinear engineered-feature evolution.** The 12-unit tanh network with 505
parameters cloned exact depth-3 five-stratum phase-safety actions and had to
reach 85% of the paired behaviour score, at least 250,000 points, and no loss
of mean moves before evolution was allowed to start. It fitted **96.86% of 255
training actions but only 43.40% of 235 disjoint teacher actions**. On the
first paired four-game screen the exact behaviour policy scored **307,222.75
points over 91.25 moves** and the nonlinear policy **148,349.25 over 47.5**;
clear and reveal throughput fell from 1.956/1.063 to 1.700/0.932 per move, and
the paired lower bounds were −441,770 points and −121.882 moves. The funnel
stopped before evolution or confirmation. Runtime 30.4 s, peak RSS 3.81 MiB.
Only `0x3d…` cloning/holdout and `0x3e…` screen seeds were read.

**Complete-game phase-weight evolution.** Six antithetic generations increased
the magnitude of cover debt by **18.4%** and trigger/rise readiness by
**26.4%**, leaving release and exposure almost unchanged — directions that
agree with the later matched oracle-topology audit. The frozen vector then
lost its eight-game screen: the default evaluator averaged **107,709.75 points
over 75.125 moves**, the evolved vector **105,534.875 over 72.5**.
Confirmation was not run.

**Both of those runs are historical 7,000-point Sequence-scored**: they used
the 7,000-point level award of the 30-drop rules rather than the 17,000-point
award of the mode studied here, so their absolute scores are not evidence
about the million-point target. The paired comparisons, the imitation rates
and the throughput figures remain valid
([mode-specific scoring correction](/docs/research/history)). Four and eight
games are small screens.

**Linear engineered-feature evolution.** The ledger references this program as
the source of *"the audited 40 action features"* reused by the nonlinear
experiment, but **no separate result section for the linear program was
located** in the [experiment history](/docs/research/history). Its own
outcome is therefore not retained; the index's family-level rejection is what
exists. Repository-verified from the source: 40 named features with per-feature
bounds, a training range of `0x3d70…` and a probe range of `0x4d70…` that are
deliberately disjoint, defaults of 12 generations × 36 candidates × 48
training games with a 64-game probe and a 750-move cap, three chance probes
during fitting and seven for the final evaluation, and a ranking objective of
`0.65 × mean + 0.35 × median` score.

</TechnicalDetails>

## What this taught us, and what is still open

- **Fitting a policy on games is not the same as improving it.** All three
  programs improved the thing they were fitted to. The strategy catalog's
  conclusion is blunt about what follows: do not treat a higher fitting score
  from another weight sweep as a new result without a fresh, paired gate.
- **A funnel that stops early is a feature.** The nonlinear run cost 30
  seconds because it checked whether its clone generalised *before* spending
  compute on evolution. That pattern (cheap falsification first) is what the
  repository asks of every experiment.
- **Plausible directions are not a result either.** Two independent methods
  agreed that cover debt and rise readiness are undervalued, and acting on
  that agreement still made the policy worse. Global rescaling of a weight
  group is too blunt an instrument to test the hypothesis it appears to test.
- **What was ruled out** is these parameterisations, fitted this way. Nothing
  here rules out learning a leaf, but the successors that tried it, with far
  more capacity, hit
  [the same wall from the same direction](/learn/concepts/ranking-siblings).