Gardner chess
Usage
or you can directly load GardnerChess
class
Description
A board needs to be five squares wide to contain all kinds of chess pieces on the first row. In 1969, Martin Gardner suggested a chess variant on 5×5 board in which all chess moves, including pawn double-move, en-passant capture as well as castling can be made. Later AISE (Associazione Italiana Scacchi Eterodossi, "Italian Heterodox Chess Association") abandoned pawn double-move and castling. The game was largely played in Italy (including by correspondence) and opening theory was developed.
Pgx implementation does not support pawn double-move, en-passant and castling.
Specs
Name | Value |
---|---|
Version | v0 |
Number of players | 2 |
Number of actions | 1225 |
Observation shape | (5, 5, 115) |
Observation type | float |
Rewards | {-1, 0, 1} |
Observation
We follow the observation design of AlphaZero [Silver+18]
.
P1 denotes the current player, and P2 denotes the opponent.
Index | Description |
---|---|
[:, :, 0:6] |
P1 board @ 0-steps before |
[:, :, 6:12] |
P2 board @ 0-steps before |
[:, :, 12:14] |
Repetitions @ 0-steps before |
... | (@ 1-7 steps before) |
[:, :, 112] |
Color |
[:, :, 113] |
Total move count |
[:, :, 114] |
No progress count |
Action
We also follow the action design of AlphaZero [Silver+18]
.
There are 1225 = 25 x 49
possible actions.
Each action represents
- 25 source position (
action // 49
), and - 49 moves (
action % 49
)
Moves are defined by 32 queen moves, 8 knight moves, and 9 underpromotions.
Rewards
Non-zero rewards are given only at the terminal states. The reward at terminal state is described in this table:
Reward | |
---|---|
Win | +1 |
Lose | -1 |
Draw | 0 |
Termination
Termination occurs when one of the following conditions are satisfied:
- checkmate
- stalemate
- no sufficient pieces to checkmate
- threefold repetition
50
halfmoves are elapsed without any captures or pawn moves256
steps are elapsed (512
in full-size chess experiments in AlphaZero[Silver+18]
)
Version History
v0
: Initial release (v1.0.0)
Baseline models
Pgx offers a baseline model for Gardner Chess. Users can use it for an anchor opponent in evaluation. See our paper for more details. See this colab for how to use it.
Model ID | Description |
---|---|
gardner_chess_v0 |
See our paper for the training details. |
Reference
[Silver+18]
"A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play" Science