← Goro

Under the hood

How Goro’s engine works

Goro does not use somebody else’s engine. The opponent you play is made for this app: written from scratch in Rust, with a sense of the game that it taught itself by playing millions of games against itself. It runs on your phone, tablet or browser, with no internet connection.

This page explains what it can do, how it learned, what makes it strong, and what we tried that did not work.

At a glance

BuiltFrom scratch in Rust. No code from any other Othello engine.
Where it runsAndroid, iOS and the web (WebAssembly), fully offline.
Judging a positionA pattern-based evaluation learned from self-play. It is a fast, compact model, not a large neural network.
SearchAlpha-beta (principal variation search) with a transposition table and ProbCut pruning.
EndgamePerfect: it solves the last 20 empty squares exactly.
Opening bookStarts from WTHOR, the French Othello Federation’s database of 137,548 tournament games (1977 to 2025), then goes beyond it: every position is scored by deep search and checked against Goro’s own engine, and the book is extended into lines humans rarely play.
SpeedAbout 4 to 8 million positions per second per core on a PC. Phones vary, some faster than the PC we measured on and some slower, so every level has its own time budget.
StrengthLevel 25 rates 2285 ± 42 on our arena Elo scale (see below).
TrainingAbout 1.5 million self-play games, about 65 million labelled positions, 39 generations.
Levels25, from a gentle beginner to full strength, each one a measured step.

How strong is it?

Goro’s strength is measured, not guessed. In a private arena, engines play each other thousands of times, and the results are turned into Elo ratings, the same kind of rating used in chess and by Othello federations: players with equal ratings score about half the games, and a player 100 points higher wins about two in three. We fit the ratings to the results by maximum likelihood and show 95% error bars.

The scale is anchored so that Edax, a well-known, very strong open-source engine, at its level 10, is worth exactly 2000. We also tested Goro against Egaroucid, another leading open-source engine. At full strength, level 25 rates 2285 ± 42, above the Edax and Egaroucid settings we tested: Edax at level 12 (2140) and Egaroucid at level 10 (2029). A rating of 2000 on this scale is a strength very few players can reach, which is why the top levels are described as approximately Othello grandmaster level.

Level 19Tournament player
1965 ± 35
Level 21Tournament player
2171 ± 37
Level 256 s per move, opening book
2285 ± 42
Dashed line: the 2000 anchor. Dots are the measured rating, the lines are the 95% error range.

Want to see how other Othello apps measure up on the same scale? Othello apps for iPhone compared places seven popular apps against Goro’s levels.

Please read these numbers with care:

  • They are engine-against-engine Elo from our own arena. They place each level against the others and against the reference engine; they are not federation ratings.
  • Level 25 gets its strength from search time, the opening book and perfect endgames, not from one big trick. The levels were calibrated on a high-end PC, using a single core. Phones vary: many recent ones are as fast as that PC or faster, and others are slower, so the same level may search a little deeper or less deep depending on the device.

How it learned

Goro was not given a library of expert knowledge. It started with a rough guess and improved itself, one generation at a time, in a loop:

  1. Start with a guessGeneration 0 uses simple, hand-made rules: corners are good, giving away corners is bad.
  2. Play against itselfEach generation plays about 40,000 games against itself. The first 8 to 20 moves are random for variety, then the engine plays with a little controlled noise.
  3. Label every positionInstead of waiting to see who won, each position is labelled with what a deep search says the final disc margin will be. Positions with 20 or fewer empty squares get the exact, solved result.
  4. Train the evaluationA model that scores positions from board patterns (edges, corners, diagonals and so on) is trained on a GPU with the newest 20 generations of data. The learning rate is lowered whenever progress stalls.
  5. Gate itThe new generation plays 800 games against the current best. It is promoted only if it is better at 95% confidence. Most are rejected: of the first 32 generations, only 10 were promoted.
  6. Rate itEach promoted generation is also rated in the arena against strong reference engines, so progress is always measured the same way.

Then the new best engine plays the next round of games, back to step 2.

A rejected generation is not wasted: it still counts as games played, and as knowledge about what does not work.

How many games

WhatGames
Self-play for training, generations 1 to 39 (about 40,000 each, kept or rejected)about 1,500,000
Arena rating matches against reference engines and our own variantsabout 105,000
Gate matches deciding whether a generation is promoted (generations 1 to 32)about 16,000
Training-recipe experiments (11 variants tested at three search depths)about 26,000
Win/draw/loss endgame experimentabout 5,400
Total, accepted and discardedabout 1.65 million

The labelled positions number about 65 million.

What makes it strong

These are the levers we turn, roughly in order of how much they mattered.

  1. Search depth and speed. The biggest lever. A faster engine searches deeper in the same time. We use transposition tables, good move ordering and an endgame solver that is always exact, though about 4 times slower than a leading open-source solver on a hard benchmark.
  2. ProbCut pruning. Many lines can be skipped when a shallow search already shows they will not matter. This lets level 25 look a few plies deeper in the same 6 seconds. It is off for the lower levels, so their calibrated strength does not change.
  3. Better training labels. When generations stopped improving at depth 8, relabelling the positions at depth 14 with ProbCut (and exact results from 20 empties) restarted progress.
  4. The training recipe. A decaying learning rate and the right amount of data (the newest 20 generations, not all of them) lowered the validation error by about 7% on the same data, and produced generation 32.
  5. A perfect endgame. From 20 empty squares, the engine plays perfectly. Gaining a few more empties here is worth little, because it is already deep.
  6. An opening book. It starts from WTHOR tournament games, then goes beyond them: each position is analysed by deep search, checked against Goro’s own engine, and extended into lines that humans rarely play. At the top level Goro varies its choices, so games do not repeat.
  7. Using time well. A new search depth is only started if it can finish. Stopping the wasted work on unfinished searches recovered about half of the thinking time.

Made for learning: 25 steps

A strong engine is no fun if it only crushes you. Goro has 25 levels, and each one is a measured step up from the one before. The weakest levels play recognisable human habits instead of random mistakes. Every level is rated, so you can watch your own progress as you beat higher ones. The top level is a serious test, and the hint system uses the same engine as the top level. See the levels on the Goro page.

What we tried that did not pan out

We keep a written log of every decision, including the failures. These are the main ones.

IdeaWhat happenedWhy we dropped it
Win / draw / loss endgame instead of exact score (a “level 26”)4,800 games at equal time: +5.3 Elo, 99% interval ± 12.3Level 25 already solves the endgame deep with ProbCut, so there is little left to gain. It also stops choosing the biggest winning margin. It is below our +20 Elo target, so it stays out of the app.
Bigger evaluation: a neural network or more patternsBetter per position (up to +75 Elo at equal nodes), but worse at equal timeThe network evaluates much more slowly, so the engine searches far less. A better evaluation that is too slow loses.
Deeper labels from the start (depth 10, exact from 20)About 3 times slower per generationNo measurable gain over depth 8 at the time. Relabelling later, with ProbCut, did work.
Training from scratch, stronger regularisation, using all old dataNo improvementOld data comes from much weaker versions of the engine and teaches the model less.
More memory for the search (64 MB instead of 32 MB)45% score: no gainSearch at this speed is not limited by table size.
A bigger opening bookExpected at most about +10 Elo, not measuredLittle to gain for a lot of computing time.
Many training generations (24 to 31, then 33 to 37)Rejected by the gateThe model had learned what its own depth-8 teacher could show it. More games did not help until the labels got deeper.

Two practical lessons. First, every change must be tested at equal time, not equal nodes: a slower idea always looks good on paper. Second, the arena gate is strict on purpose: with 800 games it can only confirm real gains of about 25 Elo or more, so small changes are not accepted by accident.

Numbers on this page come from our own arena (Edax level 10 = 2000), measured in November 2026 with generation 32, and from the project’s decision log. Edax and Egaroucid, probably the strongest Othello engines ever built, are only ever run as separate test opponents; no code from them is used in Goro.