PGX

EnvPool includes native C++ implementations of PGX environments from PGX 2.6.0.

Supported Tasks

EnvPool registers the following PGX tasks:

Go9x9-v1
Go13x13-v1
Go19x19-v1
ChineseGo9x9-v1
ChineseGo13x13-v1
ChineseGo19x19-v1
TicTacToe-v1
ConnectFour-v1
Hex-v1
Othello-v1
KuhnPoker-v1
LeducHoldem-v1
Play2048-v1
AnimalShogi-v1
Backgammon-v1
Chess-v1
GardnerChess-v1
Shogi-v1
SparrowMahjong-v1

The PGX MinAtar tasks are not registered here because EnvPool already provides native Atari environments.

Go Rules

Go9x9-v1, Go13x13-v1, and Go19x19-v1 follow PGX’s Go v1 rules:

Tromp-Taylor scoring.
N * N + 1 discrete actions, where the final action is pass.
Boolean observation shape (N, N, 17) using AlphaGo Zero history planes.
SSK legal-action filtering, with positional superko occurrence ending the game as a loss for the player who made the repeated position.

EnvPool also provides Chinese-rule variants:

Chinese area scoring: stones plus empty regions bordered by exactly one color, with neutral empty regions counted for neither player.
Positional superko moves are masked as illegal actions instead of being accepted and then turned into a terminal loss.
The same no-suicide, two-pass terminal, action, and observation API as the PGX-compatible Go tasks.

API Notes

PGX turn-based games are exposed through EnvPool’s multiplayer API. Each state contains one observation per player, info["current_player"] identifies the player ID whose turn it is, and each environment consumes one action for that current player.

The task IDs intentionally follow EnvPool style and do not use a PGX prefix.

Configuration

Go tasks support komi, history_length, max_terminal_steps, and rules. komi defaults to 7.5 and history_length defaults to 8 to match PGX. max_terminal_steps=0 means 2 * N * N. rules is "pgx" for the PGX-compatible tasks and "chinese" for the Chinese-rule variants.