File size: 12,265 Bytes
7aca0bc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Chess Environment for OpenEnv using moonfish.
This module provides a chess environment for reinforcement learning,
using python-chess for game logic and moonfish for position evaluation
and opponent play.
"""
import random
import threading
import uuid
from typing import Any, Dict, List, Optional, Tuple
import chess
from moonfish.lib import search_move
from moonfish.psqt import board_evaluation, BOARD_EVALUATION_CACHE
from openenv.core.env_server import Environment
from ..models import ChessAction, ChessObservation, ChessState
from .rubrics import ChessWinLossRubric
# moonfish keeps one evaluation cache per process, so sessions take turns with it.
_MOONFISH_LOCK = threading.Lock()
class ChessEnvironment(Environment):
"""
Chess environment implementing the OpenEnv interface.
Uses python-chess for game logic and moonfish for position evaluation.
Designed for RL training where an agent plays as one color against
an opponent (which can be random, moonfish engine, or self-play).
"""
SUPPORTS_CONCURRENT_SESSIONS = True
def __init__(
self,
opponent: str = "moonfish",
opponent_depth: int = 2,
max_moves: int = 500,
agent_color: str | None = None,
gamma: float = 0.99,
):
"""
Initialize the chess environment.
Args:
opponent: Opponent type - "moonfish", "random", or None (self-play)
opponent_depth: Search depth when using moonfish as opponent
max_moves: Maximum half-moves before draw
agent_color: Which color the agent plays - "white", "black", or None (alternate each episode)
gamma: Discount factor for temporal credit assignment (0-1)
"""
super().__init__(rubric=ChessWinLossRubric(gamma=gamma))
self._opponent = opponent
self._opponent_depth = opponent_depth
self._max_moves = max_moves
self._agent_color_setting = agent_color
self._gamma = gamma
self._board = None
self._state = None
self._agent_color = chess.WHITE
self._agent_move_count = 0 # Track agent moves for discounting
self.reset()
def reset(self, fen: str = None):
"""
Initialize a new chess game episode.
Args:
fen: Optional starting position in FEN notation.
Returns:
Initial observation of the board state.
"""
self._reset_rubric()
# moonfish's module-level evaluation cache is never trimmed.
with _MOONFISH_LOCK:
BOARD_EVALUATION_CACHE.clear()
if fen:
self._board = chess.Board(fen)
else:
self._board = chess.Board()
# Generate episode ID once
episode_id = str(uuid.uuid4())
# Determine agent color
if self._agent_color_setting == "white":
self._agent_color = chess.WHITE
elif self._agent_color_setting == "black":
self._agent_color = chess.BLACK
elif self._agent_color_setting is None:
# Alternate each episode based on episode_id hash
self._agent_color = (
chess.WHITE if hash(episode_id) % 2 == 0 else chess.BLACK
)
else:
self._agent_color = chess.WHITE
self._state = ChessState(
episode_id=episode_id,
step_count=0,
current_player="white" if self._board.turn else "black",
fen=self._board.fen(),
move_history=[],
)
self._agent_move_count = 0
# If agent plays Black and opponent is configured, opponent moves first
if self._opponent is not None and self._agent_color == chess.BLACK:
self._make_opponent_move()
# Check if starting position is already terminal (e.g., custom FEN with checkmate/stalemate)
_, done = self._calculate_reward_and_done()
return self._make_observation(done=done)
def step(self, action: ChessAction):
"""
Execute a chess move and return the resulting state.
Args:
action: The move to make in UCI format (e.g., "e2e4").
Returns:
Observation with reward and done flag.
"""
if self._board is None:
raise RuntimeError("Environment not initialized. Call reset() first.")
# Parse the move
try:
move = chess.Move.from_uci(action.move)
except ValueError:
return self._make_observation(reward=-0.1, done=False)
# Check if move is legal
if move not in self._board.legal_moves:
return self._make_observation(reward=-0.1, done=False)
# Execute the move
self._board.push(move)
self._state.step_count += 1
self._agent_move_count += 1
self._state.move_history.append(action.move)
self._state.current_player = "white" if self._board.turn else "black"
self._state.fen = self._board.fen()
# Calculate reward and check for game end
reward, done = self._calculate_reward_and_done()
# If game not over and opponent is configured, make opponent move
if not done and self._opponent is not None:
self._make_opponent_move()
reward, done = self._calculate_reward_and_done()
obs = self._make_observation(reward=reward, done=done)
# Feed to rubric for trajectory accumulation
self._apply_rubric(action, obs)
# Note: obs.reward is kept from _calculate_reward_and_done() rather than
# overridden by the rubric. The rubric here is a TrajectoryRubric that
# returns 0.0 until done then the final score — the inline reward logic
# already provides the correct per-step reward including penalties for
# invalid/illegal moves. The rubric's value is in compute_step_rewards()
# for training infrastructure, not in replacing obs.reward.
return obs
def _make_observation(self, reward: float = 0.0, done: bool = False):
"""Build observation from current board state."""
legal_moves = [move.uci() for move in self._board.legal_moves]
result = None
if done:
result = self._get_result_string()
with _MOONFISH_LOCK:
evaluation = board_evaluation(self._board)
metadata = {
"evaluation": evaluation,
"fullmove_number": self._board.fullmove_number,
"halfmove_clock": self._board.halfmove_clock,
}
# Compute discounted rewards for all agent moves when episode ends
if done and self._agent_move_count > 0:
discounted_rewards = self._compute_discounted_rewards(reward)
metadata["discounted_rewards"] = discounted_rewards
metadata["gamma"] = self._gamma
return ChessObservation(
fen=self._board.fen(),
legal_moves=legal_moves,
is_check=self._board.is_check(),
done=done,
reward=reward,
result=result,
metadata=metadata,
)
def _calculate_reward_and_done(self):
"""Calculate reward and check if episode is done."""
if self._board.is_checkmate():
winner = not self._board.turn
if winner == self._agent_color:
return 1.0, True
else:
return -1.0, True
if self._board.is_stalemate():
return 0.0, True
if self._board.is_insufficient_material():
return 0.0, True
if self._board.is_fifty_moves():
return 0.0, True
if self._board.is_repetition(3):
return 0.0, True
if self._state.step_count >= self._max_moves:
return 0.0, True
return 0.0, False
def _get_result_string(self):
"""Get the game result as a string."""
if self._board.is_checkmate():
return "1-0" if not self._board.turn else "0-1"
return "1/2-1/2"
def _compute_discounted_rewards(self, terminal_reward: float) -> list[float]:
"""
Compute temporally discounted rewards for all agent moves.
Uses exponential discounting: r_t = γ^(T-1-t) * R_final
where T is total agent moves, t is move index, R_final is terminal reward.
Earlier moves get less credit, later moves get more credit.
This helps with credit assignment in long games.
Args:
terminal_reward: The final reward (+1 win, -1 loss, 0 draw)
Returns:
List of discounted rewards, one per agent move
"""
T = self._agent_move_count
discounted_rewards = []
for t in range(T):
# γ^(T-1-t) means last move (t=T-1) gets γ^0 = 1.0
# First move (t=0) gets γ^(T-1)
discount = self._gamma ** (T - 1 - t)
discounted_rewards.append(discount * terminal_reward)
return discounted_rewards
def _make_opponent_move(self):
"""Make a move for the opponent using configured strategy."""
if not self._board.legal_moves:
return
if self._opponent == "moonfish":
with _MOONFISH_LOCK:
move = search_move(self._board, depth=self._opponent_depth)
elif self._opponent == "random":
move = random.choice(list(self._board.legal_moves))
else:
return
self._board.push(move)
self._state.step_count += 1
self._state.move_history.append(move.uci())
self._state.current_player = "white" if self._board.turn else "black"
self._state.fen = self._board.fen()
def web_actions(
self, observation: Dict[str, Any]
) -> List[Tuple[str, Dict[str, Any]]]:
"""The legal moves (UCI) as buttons, sorted so moves of the same piece sit together."""
return [
(move, {"move": move})
for move in sorted(observation.get("legal_moves") or [])
]
def render_web(self, observation: Dict[str, Any]) -> Optional[str]:
"""Draw the board from the FEN, white at the bottom, the king in check highlighted."""
if "fen" not in observation:
return None
board = chess.Board(observation["fen"])
checked = board.king(board.turn) if observation["is_check"] else None
cells = []
for rank in range(7, -1, -1):
cells.append(f'<span style="opacity:.6;font-size:12px">{rank + 1}</span>')
for file in range(8):
square = chess.square(file, rank)
piece = board.piece_at(square)
# Fixed board colours, so a1 stays dark in both themes.
if square == checked:
bg = "var(--color-accent)"
elif (file + rank) % 2:
bg = "hsl(37 50% 84%)"
else:
bg = "hsl(28 33% 55%)"
# U+FE0E keeps the pawn as a text glyph instead of an emoji.
glyph = piece.unicode_symbol() + "\ufe0e" if piece else ""
cells.append(
f'<span style="background:{bg};color:black">{glyph}</span>'
)
cells.append("<span></span>")
cells += [
f'<span style="opacity:.6;font-size:12px;line-height:20px">{f}</span>'
for f in "abcdefgh"
]
return (
'<div role="img" aria-label="Chess board" style="display:inline-grid;'
"grid-template-columns:16px repeat(8,40px);grid-template-rows:repeat(8,40px) 20px;"
"line-height:40px;text-align:center;font-size:30px;color:var(--body-text-color);padding:10px;"
'border:1px solid var(--border-color-primary);border-radius:10px;background:var(--background-fill-secondary)">'
+ "".join(cells)
+ "</div>"
)
@property
def state(self):
"""Return the current episode state."""
return self._state
|