Marble Solitaire AI

Watch an AlphaZero-style neural network learn to solve European peg solitaire

How It Works

The Game

37 holes, 36 marbles, one empty spot in the middle. Jump a marble orthogonally (no diagonals) over a neighbor into an empty hole; the jumped marble is removed. The goal: end with as few marbles as possible.

The AI

A small neural network (75K parameters) trained itself by playing thousands of games against itself — no rules of strategy taught, just the legal moves. Same recipe as AlphaGo, scaled way down.

Watch the model learn

At: Clueless

Marbles left: 12

The dot is the same neural network getting better with training. It plateaus at 2 — the proven mathematical minimum from this start.

The 5 Generations

Each generation is the same network at a later point in training. Pick one and hit Solve to see how it plays:

Gen 1 — Clueless: random play, ~12 marbles left.
Gen 2 — Beginner: picking up patterns, ~8 left.
Gen 3 — Intermediate: real strategy, ~8 left.
Gen 4 — Advanced: strong play, ~5 left.
Gen 5 — Master: 2 marbles — the proven optimum.

Why Gen 5 stops at 2

From a centre-start on the 37-hole European board, no single-marble finish is reachable. A computer enumeration by Durango Bill checked all 16,662,591,542,028,595,169,821,912 possible game sequences from this start — that's 16 quintillion — and zero end with one marble. The mathematical minimum is 2, and Gen 5 hits it every game.

Or switch to Manual and try it yourself.

Generation: → → → →

Marbles 36

Move 0

Confidence --

Status Ready

Speed