Monte Carlo Tree Search

Copyright (c) 2021 Daniel Kanczuk

Easy to implement alghoritm creating a game with interface:

type  State  interface {
	Simulate() SimulationResult
	Expand(Iteration) State
	Iterations() []Iteration
	Copy() State
	ID() string
}

Simulate: Play a random game based on current state.

Iterations: Total iterations of current state.

Expand: Node expansion based on an previews Iteration sent

Copy: Deep copy of State. When implementing this interface, take care of copying objets in a way that doesn't carry any reference from the other one

ID: State identifier.It's a way to prevent duplicated instances

An example of implement:

game := yourCustomGame{}
tree := mcts.NewMonteCarloTree(mcts.MonteCarloTreeConfig{MaxIterations: 256})

// the command start will block until process end
node, err := tree.Start(game)
if err != nil {
    panic(err)
}
// the new state of game, is the result of tree processing
newGameState := node.NodeScore[0].State.(yourCustomGame)

// you can use the newGameState again in monte carlo tree
// inner loop may be necessary to do it!

This implementation is based on example of tic-tac-toe AI and it was implemented with one expasion per iteration instead of expand all, to prevent memory overhead.

Check some examples here:

• tic tac toe
• game 2048