MCTS (Tree Search)

MCTS explores possible future moves by trying many pretend games.

Entry 42 / 58Try-it-yourself page

What it does

MCTS explores possible future moves by trying many pretend games.

Simulation: Mental practice game • Roll-out: Playing to the end

How it works

  • Pick a possible move.
  • Simulate what might happen after that move.
  • Choose moves that looked good in many simulations.

Uses & Applications

  • A board game AI can choose a smart move.
  • A planning tool can compare possible next steps.
  • A robot can test possible paths in a simulation before moving.
In short

MCTS is closely linked to AlphaGo and game search.

Try it yourself

Try future moves before choosing

In this stone game, take one or two stones. The last person to take a stone wins. The computer tries possible future games before choosing.

Your turn

Change one setting at a time. Reset example puts everything back.

Search a take-away game tree

Step 1
Why this happens

Look closer at the numbers
Numbers used in this example

Related methods

Back to Reinforcement learning, agents and planningAll topics