Players

AbstractPlayer{Game}

Abstract type for a player of Game.

think(::AbstractPlayer, state, turn=nothing)

Return a probability distribution over actions as a (actions, π) pair.

The turn argument, if provided, indicates the number of actions that have been played before by both players in the current game. It is useful as during self-play, AlphaZero typically drops its temperature parameter after a fixed number of turns.

select_move(player::AbstractPlayer, state, turn=nothing)

Return a single action. A default implementation is provided that samples an action according to the distribution computed by think.

reset_player!(::AbstractPlayer)

Reset the internal memory of a player (e.g. the MCTS tree). The default implementation does nothing.

MctsPlayer{Game, MctsEnv} <: AbstractPlayer{Game}

A player that selects actions using MCTS.

Constructors

MctsPlayer(mcts::MCTS.Env; τ, niters, timeout=nothing)

Construct a player from an MCTS environment. When computing each move:

• if timeout is provided, MCTS simulations are executed for timeout seconds by groups of niters
• otherwise, niters MCTS simulations are run

The temperature parameter τ can be either a real number or a StepSchedule.

MctsPlayer(oracle::MCTS.Oracle, params::MctsParams; timeout=nothing)

Construct an MCTS player from an oracle and an MctsParams structure. If the oracle is a network, this constructor handles copying it, putting it in test mode and copying it on the GPU (if necessary).

RandomPlayer{Game} <: AbstractPlayer{Game}

A player that picks actions uniformly at random.

NetworkPlayer{Game, Net} <: AbstractPlayer{Game}

A player that uses the policy output by a neural network directly, instead of relying on MCTS.

EpsilonGreedyPlayer{Game, Player} <: AbstractPlayer{Game}

A wrapper on a player that makes it choose a random move with a fixed $ϵ$ probability.

play_game(white, black, memory=nothing)

Play a game between two AbstractPlayer and return the reward obtained by white.

• If the memory argument is provided, samples are automatically collected from this game in the given MemoryBuffer.
• If the flip_probability argument is set to $p$, the board is flipped randomly at every turn with probability $p$, using GI.random_symmetric_state.

pit(handler, contender, baseline, ngames)

Evaluate two AbstractPlayer against each other in a series of games.

Arguments

• handler: this function is called after each simulated game with two arguments: the game number i and the collected reward z for the contender player
• ngames: number of games to play

Optional keyword arguments

• reset_every: if set, players are reset every reset_every games
• color_policy: determines the ColorPolicy, which is ALTERNATE_COLORS by default
• memory=nothing: memory to use to record samples
• flip_probability=0.: see play_game

@enum ColorPolicy ALTERNATE_COLORS BASELINE_WHITE CONTENDER_WHITE

Policy for attributing colors in a duel between a baseline and a contender.

interactive!(game, white, black)

Launch an interactive session for game::AbstractGame between players white and black. Both players have type AbstractPlayer and one of them is typically Human.

Human{Game} <: AbstractPlayer{Game}

Human player that queries the standard input for actions.

Does not implement think but instead implements select_move directly.