Every other family on this shelf searches from first principles: a backtracker
digs, a beam builds, a local search polishes, each reasoning only from the rules
of the puzzle and the board in front of it. This family does something different.
It takes the corpus of strong boards people have already found, reads it for
structure, and feeds that structure back into a search as a bias. It is closer to
imitation learning than to search design: the question is not "what is a good
move" from first principles but "what did the boards that scored well tend to do,
and can that be reused."
The pages below go technique by technique. Each is genuinely its own method, so
they do not share a shape the way the local-search polishers do: a position prior,
a learned move-ranking, an anti-pattern miner, and a witness replay share a thesis
but not a mechanism. For the whole territory at once, start with
a map of every known approach. For the runnable
experiments that embody these techniques, with their boards and numbers, see the
learning-from-boards experiments.
A learned signal helps when it captures real, transferable structure: something
true of good boards that a from-scratch search would otherwise have to rediscover
by luck. Where each piece tends to sit, which pieces like to touch, which scarce
demand a rare piece is the only one that can serve. Wired into a search as a gentle
bias, that structure reliably carries a build to the top of its own range.
The limit is the part that makes this research rather than a bag of tricks. The
corpus these signals are learned from is itself sub-perfect: every board in it is
stuck somewhere below 480. So a signal mined from the corpus encodes the
community's ceiling as much as its wisdom, and a signal that merely re-encodes
the plateau every strong board is already caught on raises nothing. Push a learned
prior from a tiebreak into the objective and the search collapses, because it starts
trading the edge in front of it against a statistical hunch. The plain through-line
of this whole family: learning from strong boards helps you reach the plateau fast
and reliably, but it does not, on its own, get you past it. The ceiling is the
rigidity wall, and no amount of corpus mining moves it.
- Corpus priors are the simplest
form: count where each piece sits in strong boards, or how often a piece serves a
scarce demand, and use the count as a tiebreak in construction. The
PRIOR and
LODESTONE experiments.
- Learned move-ordering
ranks candidate placements by several learned signals at once, and lets them vote,
so no single hunch marches the search into the same dead end. The
KEYRING experiment.
- Anti-pattern mining is the
subtler idea: not every agreement across strong boards is good. Separate the real
structure from the shared trap, and attack the trap. The
PALIMPSEST experiment,
which reached this project's best board.
- Decoding records reconstructs a
known record board exactly, piece for piece, to recover the ingredient an ordinary
search was missing. Reproduction as proof. The
REPLAY experiment.
- When learning collapses is
the failure-mode capstone: exactly when a learned signal stops helping, why
over-trusting it wrecks the search, and why none of these methods raise the ceiling.
It sets the limit for the whole section: learned signals reach the plateau
faster, but none of them lift it.