Monte Carlo Tree Search — ELI5

Imagine you are at a fork in a hiking trail. One path goes left, the other right. You have no map. How do you pick?

One idea: send a hundred imaginary scouts down each path. They walk randomly, sometimes hitting dead ends, sometimes finding a beautiful lake. You count how many scouts from each path found something good, and you take the path with more happy scouts.

That is exactly how Monte Carlo Tree Search (MCTS) works. A computer program looks at a game board and imagines thousands of quick, random games from that position. Each random game is like sending one scout down one path. After many scouts, the program counts which first move led to the most wins and picks that one.

The trick is that MCTS does not just pick randomly every time. It remembers which branches worked well before and sends more scouts down promising paths, while still exploring new ones occasionally. This balance between trying known good moves and exploring unknown ones is the secret sauce.

MCTS is famous because it powered AlphaGo, the first program to beat a world champion at Go — a game so complex that no computer could brute-force all the moves. Instead of trying every possibility, MCTS sampled the most interesting ones and made surprisingly good decisions.

The beauty is that MCTS needs no special knowledge about the game. Give it the rules and enough random playouts, and it figures out strategy on its own.

The one thing to remember: MCTS picks the best move by playing thousands of quick imaginary games and trusting the move that wins most often.