User Interface

The default user interface for AlphaZero.

The user interface is fully separated from the core algorithm and can therefore be replaced easily.

Session{Env}

A wrapper on an AlphaZero environment that adds features such as:

• Logging and plotting
• Loading and saving of environments

In particular, it implements the Handlers interface.

Public fields

• env::Env is the environment wrapped by the session
• report is the current session report, with type SessionReport

Session(::Type{Game}, ::Type{Net}, params, netparams) where {Game, Net}

Create a new session using the given parameters, or load it from disk if it already exists.

Arguments

• Game is the type ot the game that is being learnt
• Net is the type of the network that is being used
• params has type Params
• netparams has type Network.HyperParams(Net)

Optional keyword arguments

• dir: session directory in which all files and reports are saved; this argument is either a string or nothing (default), in which case the session won't be saved automatically and no file will be generated
• autosave=true: if set to false, the session won't be saved automatically nor any file will be generated
• nostdout=false: disables logging on the standard output when set to true
• benchmark=[]: vector of Benchmark.Duel to be used as a benchmark
• load_saved_params=false: if set to true, load the training parameters from the session directory (if present) rather than using the params argument
• save_intermediate=false: if set to true (along with autosave), all intermediate training environments are saved on disk so that the whole training process can be analyzed later. This can consume a lot of disk space.

Session(::Type{Game}, ::Type{Network}, dir::String) where {Game, Net}

Load an existing session from a directory.

This constructor accepts the optional keyword arguments autosave, nostdout, benchmark and save_intermediate.

Session(env::Env[, dir])

Create a session from an initial environment.

• The iteration counter of the environment must be equal to 0
• If a session directory is provided, this directory must not exist yet

This constructor features the optional keyword arguments autosave, nostdout, benchmark and save_intermediate.

resume!(session::Session)

Resume a previously created or loaded session. The user can interrupt training by sending a SIGKILL signal.

save(session::Session)

Save a session on disk.

This function is called automatically by resume! after each training iteration if the session was created with autosave=true.

play_interactive_game(session::Session; timeout=2.)

Start an interactive game against AlphaZero, allowing it timeout seconds of thinking time for each move.

start_explorer(session::Session)

Start an explorer session for the current environment. See Explorer.

SessionReport

The full collection of statistics and benchmark results collected during a training session.

Fields

• iterations: vector of $n$ iteration reports with type Report.Iteration
• benchmark: vector of $n+1$ benchmark reports with type Benchmark.Report

Explorer{Game}

A command interpreter to explore the internals of a player through interactive play.

Constructors

Explorer(player::AbstractPlayer, game=nothing; memory=nothing)

Build an explorer to investigate the behavior of player in a given game (by default, in the initial state). Optionally, a reference to a memory buffer can be provided, in which case additional state statistics will be displayed.

Explorer(env::Env, game=nothing; arena_mode=false)

Build an explorer for the MCTS player based on neural network env.bestnn and on parameters env.params.self_play.mcts or env.params.arena.mcts (depending on the value of arena_mode).

Commands

The following commands are currently implemented:

• do [action]: make the current player perform action. By default, the action of highest score is played.
• explore [num_sims]: run num_sims MCTS simulations from the current state (for MCTS players only).
• go: query the user for a state description and go to this state.
• flip: flip the board according to a random symmetry.
• undo: undo the effect of the previous command.
• restart: restart the explorer.

start_explorer(exp::Explorer)

Start an interactive explorer session.