How hard is a puzzle level? Tile counts are a poor answer. A 144-tile layout can be gentler than a 100-tile one if the shape is forgiving. When I added 12 layouts to Instant Mahjong Solitaire, I wanted a number that tells players and me something real.
This post is about one cheap trick: measuring difficulty with random play.
The layouts
The game has 12 layouts, from 60 to 144 tiles: Harmony, Garden, Phoenix, Pyramid, Pagoda, Crane, Lotus, Butterfly, Dragon, Turtle, Bridge and Mountain. They differ in height, symmetry and how many tiles start locked under others.
Every deal is generated so that a winning path exists. That removes "impossible" from the equation, which leaves a different question: how easy is it to find the path?
The random-play bot
The measuring tool is almost embarrassingly simple.
// simplified sketch
function randomPlay(board, rng) {
while (true) {
const moves = freePairs(board);
if (moves.length === 0) return board.isEmpty(); // cleared or stuck
board.remove(moves[Math.floor(rng() * moves.length)]);
}
}
function clearRate(layout, runs = 2000) {
let wins = 0;
for (let i = 0; i < runs; i++) {
if (randomPlay(deal(layout, seed(i)), mulberry(i))) wins++;
}
return wins / runs;
}
The bot picks a random legal pair every turn, with no hints and no lookahead. If it clears a layout often, careless play is enough to win, so the layout is gentle. If it almost never clears, the player must think.
Reading the numbers
On the layouts I tested, clear rates ranged widely. Garden sat around 70%, and Mountain around 20%. The numbers are a rough guide, not a law, but they give three benefits:
- Ordering. I can sort layouts from calm to demanding based on data and not my own feeling.
- Surprises. A layout I thought was hard but that the bot clears easily is a sign the shape is too open.
- Regression checks. If I change the generator or the matching rules, I rerun the numbers and see whether difficulty drifted.
What random play does not capture
Be honest about limits:
- A random bot does not use unlimited undo and hints, which human players have. So human clear rates are higher than the bot's.
- The bot does not plan. A skilled human can clear boards the bot nearly never does.
- It measures the chance of stumbling into a win, not the enjoyment of the layout.
So I treat the clear rate as a floor for difficulty, not the final word.
Using it for design
Once you have a number, you can aim for it. If I want a "medium" layout, I look for one in the middle of the range. If a new layout lands at 5%, it is probably too punishing for a mainstream audience and might be better as a late-game challenge.
Why this works for other puzzles
The same approach fits any puzzle with a legal-move generator: sliding puzzles, match-3 boards, block puzzles. Build a dumb bot, run it a few thousand times, and plot the results. It costs an afternoon and tells you things playtests miss.
Try the layouts
You can see how they feel in Instant Mahjong Solitaire, a free mahjong game online with no download.
Have you measured difficulty with a bot? What was your metric?
If you want mahjong solitaire tips, start with a smaller layout and work up. It is a free mahjong solitaire game with mahjong online access in any browser.
Top comments (0)