Package mcts provides parallel monte-carlo tree search for your Go applications.
Go latest is recommended.
go get -u github.com/go-mcts/mctsSee examples directory.
Implements mcts.Move and mcts.State:
// examples/nim/nim.go
type Move int
type State struct {
playerToMove int
chips int
}
func (s *State) PlayerToMove() int {
return s.playerToMove
}
func (s *State) HasMoves() bool {
s.checkInvariant()
return s.chips > 0
}
func (s *State) GetMoves() []mcts.Move {
s.checkInvariant()
var moves []mcts.Move
for i := 1; i <= min(3, s.chips); i++ {
moves = append(moves, i)
}
return moves
}
func (s *State) DoMove(move mcts.Move) {
m := move.(int)
if m < 1 || m > 3 {
panic("illegal move")
}
s.checkInvariant()
s.chips -= m
s.playerToMove = 3 - s.playerToMove
s.checkInvariant()
}
func (s *State) DoRandomMove(rd *rand.Rand) {
if s.chips <= 0 {
panic("invalid chips")
}
s.checkInvariant()
max := min(3, s.chips)
s.DoMove(rd.Intn(max) + 1)
s.checkInvariant()
}
func (s *State) GetResult(currentPlayerToMove int) float64 {
if s.chips != 0 {
panic("game not over")
}
s.checkInvariant()
if s.playerToMove == currentPlayerToMove {
return 1.0
}
return 0.0
}
func (s *State) Clone() mcts.State {
return &State{
playerToMove: s.playerToMove,
chips: s.chips,
}
}Run mcts.ComputeMove:
state := &State{
playerToMove: 1,
chips: chips,
}
move := mcts.ComputeMove(state, mcts.MaxIterations(100000), mcts.Verbose(true))By default, runtime.NumCPU() goroutines are started for compute:
2021-10-16T16:13:33+08:00 DEBUG 100000 games played (493618.37 / second).
2021-10-16T16:13:33+08:00 DEBUG 100000 games played (487598.00 / second).
2021-10-16T16:13:33+08:00 DEBUG 100000 games played (474579.62 / second).
2021-10-16T16:13:33+08:00 DEBUG 100000 games played (470514.17 / second).
2021-10-16T16:13:33+08:00 DEBUG Move: 2 (16% visits) (48% wins)
2021-10-16T16:13:33+08:00 DEBUG Move: 1 (72% visits) (67% wins)
2021-10-16T16:13:33+08:00 DEBUG Move: 3 (13% visits) (48% wins)
2021-10-16T16:13:33+08:00 DEBUG Best: 1 (72% visits) (67% wins)
2021-10-16T16:13:33+08:00 DEBUG 400000 games played in 0.21 s. (1881366.23 / second, 4 parallel jobs).
This project is under the MIT License. See the LICENSE file for the full license text.