Gomoku

Five in a row on a 15×15 board, with a computer opponent. No build step, no dependencies.

▶ Play it here

Or clone and open index.html directly — there is nothing to install and no server to run:

open index.html

Rules

Freestyle Gomoku: black moves first, and the first player to get five or more of their stones in an unbroken line — horizontal, vertical, or diagonal — wins. There are no forbidden moves, and an overline (six or more) counts as a win.

Playing

Click an intersection to place a stone. The panel tracks whose turn it is, the running score across games, and the move list in H8 notation.

Opponent	Computer as white or black, two players at one keyboard, or computer vs computer
Difficulty	Easy, Medium, Hard
Undo	Steps back to your own turn — against the computer that means taking back both plies
U / N	Undo / new game

Switching opponent mode starts a fresh game; changing difficulty applies from the computer's next move.

How the computer plays

js/ai.js is an iterative-deepening alpha–beta search over a pattern-based evaluation.

Evaluation. Each line on the board — every row, column, and diagonal — is rendered as a string from one player's point of view (1 own stone, 2 opponent or wall, 0 empty) and scored by counting known shapes in it: an open four is worth far more than a blocked one, an open three more than a dead three, and so on. Shapes are written with their surrounding context, so 011110 (open four) and 211110 (blocked four) score differently.

Incremental scoring. Placing a stone can only change the four lines through that square, so make/unmake recomputes just those and adjusts a running total. Line scores are memoised on the line's text, which recurs constantly during a search. Together these make a node cheap enough to search six plies in well under a second.

Search. Move generation is restricted to empty squares near existing stones. Ordering does real work near the root — ranking a square by what it gains for both players, since a square valuable to the opponent is exactly the one worth taking away — and nothing near the leaves, where nodes are too numerous to pay for it. Because the leaf ordering is cheap and narrow, the search checks tactics itself before ordering: it returns immediately on a move that makes five, and if the opponent threatens five it considers only blocking replies. That keeps a narrow move list from ever hiding a forced win or loss.

Difficulty sets search depth and how often the engine settles for a move that is good but not best — Easy also skips the "block a four" shortcut, so it can be caught out.

Layout

index.html      markup
css/style.css   styling
js/board.js     board state, move history, win detection
js/ai.js        evaluation and search
js/ui.js        canvas rendering, input, turn scheduling

js/board.js has no dependencies, and js/ai.js depends only on the board. The scripts are plain globals loaded in order, so the page works over file:// with no server.