Ilia Denisov 6a794bd9e2 scrabble: validate scoring & move generation against real GCG games
Replay 17 real tournament games (cross-tables.com, including the 700+ club) and
check, move for move, that:
  - ScorePlay reproduces the protocol score and running total exactly;
  - the move generator produces every dictionary-valid played move with that score.

The GCG parser handles coordinates (across/down), blanks (lowercase), played-through
tiles ("."), passes/exchanges, challenge bonuses, end-game rack adjustments and phony
withdrawals ("--", un-applied). Uses the committed dawg/en_sowpods.dawg.
2026-06-01 23:56:39 +02:00

scrabble-solver

A Go library that, given a dictionary, a board position and a rack, returns every legal play ranked by score, and also scores or validates arbitrary plays. The move generator is the DAWG algorithm of Appel & Jacobson, The World's Fastest Scrabble Program. It operates on compact byte-indexed inputs/outputs and is dictionary-driven via github.com/iliadenisov/dafsa.

See ALGORITHM.md for the algorithm (the single source of truth) and RESULTS.md for the DAWG-vs-GADDAG benchmark that settled the design.

Status

  • DAWG move generation (across / down / both orientations), with full tournament scoring (cross-words, premiums, all-tiles bonus) and a per-tile breakdown.
  • Public Solver: GenerateMoves (ranked), ScorePlay, ValidatePlay.
  • Rulesets: English Scrabble, Russian Scrabble, Эрудит; rules.Ruleset is fully parameterizable (board, premiums, tile values/counts, blanks, rack, bonus).
  • A GADDAG (Gordon) was implemented, benchmarked and then removed — for a scoring solver it was ~7× larger and no faster.

Layout

scrabble/        public API: Solver, Move/Play types, DAWG generator, scoring, validation
board/ rack/ rules/   board grid (+transpose), rack, rulesets (English/Russian/Эрудит)
internal/        encoding (byte conventions), wordlist, dictdawg, dict, graph
cmd/builddict/   word list -> serialized DAWG in testdata
cmd/stress/      greedy self-play benchmark of the generator
selfplay/        bag + greedy player + game loop

Setup

git submodule update --init            # the dictionaries submodule (SOWPODS, TWL06, …)
go run ./cmd/builddict                 # build testdata/sowpods.dawg (≈0.2 s, ~730 KB)

go.mod carries replace github.com/iliadenisov/dafsa => ../dafsa: the solver needs dafsa's low-level traversal Cursor (see the patch notes in ../dafsa/SCRABBLE_API.md).

Usage

rs := rules.English()
finder, _ := dict.EnglishDAWG()            // loads testdata/sowpods.dawg
s := scrabble.NewSolver(rs, finder)

b := board.New(rs.Rows, rs.Cols)           // empty board (first move)

r := rack.New(rs.Size())                   // rack "friends"
tiles, _ := rs.Alphabet.Encode("friends")
for _, t := range tiles {
    r.Add(t)
}

moves := s.GenerateMoves(b, r, scrabble.Both) // ranked, highest score first
best := moves[0]
// best.Main / best.Cross hold the words (alphabet indexes; decode via rs.Alphabet),
// best.Tiles the placed tiles (with blank flags), best.Score the total.

// Score or validate an arbitrary play (placed tiles + direction):
m, err := s.ValidatePlay(b, scrabble.Horizontal, best.Tiles)
_ = m
_ = err

Words and tiles are alphabet indexes throughout (no string wrapper); convert with the ruleset's alphabet.Indexer (Encode/Decode) when you need text.

Rulesets

rules.English(), rules.RussianScrabble(), rules.Erudit(), or build your own with rules.FromTemplate(...). For Эрудит, fold Ё→Е while preparing the dictionary with wordlist.FoldYo (the engine treats them as one letter; it is a dictionary-prep step).

Benchmark

go run ./cmd/stress -games 100   # greedy AI-vs-AI self-play; reports speed and memory

Tests

go test ./...           # unit tests + a brute-force move-generation oracle
go test ./... -short    # skips the full-dictionary game test
S
Description
No description provided
Readme 22 MiB
Languages
Go 100%