mnk-bench

Play

Pick a board and two players. Every agent runs in your browser, ported from the Python package. The exact solver plays perfectly while a board is small enough to solve; on bigger boards it refuses and falls back to alpha-beta, just as in the benchmark.

Score this session

0
X wins
0
Draws
0
O wins

Custom board

Is this board solvable?

Moves

How to play

What to expect from each opponent

playeron small boards (tic-tac-toe, 4×4)on big boards (Connect Four, gomoku)time per move
Random
baseline
Picks any legal move. Easy to beat.Same: you win as soon as you make a line.instant
Rule-based
baseline
Never misses a win or a block, and makes and blocks forks. Hard to beat on tic-tac-toe.Sees only one move ahead, so two threats at once (an open three in gomoku) beat it.instant
MENACE
learning: matchboxes
Trains before its first move (a short pause), then plays sensibly on tic-tac-toe. After 2,000 training games it still loses some games to strong opponents such as MCTS, because training hasn't covered every line; choose 10,000 games for stronger play.Has met almost none of the positions, so most moves are random and you'll see a warning. This is where MENACE breaks.training ~0.1–1 s once, then instant
Minimax
exact search
Perfect play up to about (4,4,4): you can at best draw tic-tac-toe, and on boards that are a first-player win, such as (3,4,3), it wins as X. It takes the fastest win and delays losses.Refuses: solving would take years. It plays depth-4 alpha-beta instead and says so in a yellow note.under 1 ms on tic-tac-toe; up to ~1 s for the first move on (4,4,4)
Alpha-beta
heuristic search
Strong at depth 3–4, though it can occasionally be tricked because it does not search to the end.The strongest agent here. It looks depth moves ahead and scores threats, so beat it with a plan longer than its horizon. Depth 1–2 is beatable; depth 4–6 is tough.a few ms on Connect Four, ~10–50 ms on gomoku at depth 4; deeper is slower
MCTS
sampling
Plays well: random playouts are informative on small boards.Weak on open boards: random playouts say little, so it misses threats. It loses to rule-based in about 10 moves on gomoku. More playouts help, slowly.~5–50 ms at 1,000 playouts, up to the 5 s setting

Agents run in a background thread, so the page stays responsive while they think. The Last agent move panel shows what the agent did: nodes searched, playouts, MENACE's matchbox, or why it refused a board.