kuhyx/testsAndMisc
1
0
Fork 0
mirror of https://github.com/kuhyx/testsAndMisc.git synced 2026-07-06 10:03:07 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

feat: remove entire looks blunderish logic

Browse Source
This commit is contained in:
Krzysztof kuhy Rudnicki 2025-08-23 19:49:50 +02:00
parent 02cdcbd17c
commit da3665f5c7
1 changed files with 168 additions and 389 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

557
PYTHON/lichess_bot/engine.py
View File

@ -311,20 +311,6 @@ class RandomEngine:
mv = self.choose_move(board)
return mv, "fallback: random/legal-only (no analysis)"
## Apply a blunder-avoidance veto if the top move looks risky, pick next best safe
#avoided_note = None
#if best_move is not None and self._looks_blunderish(board, best_move):
# avoided_note = f"avoided risky {board.san(best_move)} ({best_move.uci()})"
# for cand, _ in scores[1:]:
# if not self._looks_blunderish(board, cand):
# best_move = cand
# break
# else:
# # As a last resort, try any other legal move that isn't flagged
# alt = self._pick_safer_alternative(board, avoid=best_move)
# if alt is not None:
# best_move = alt
# Build explanation
def annotate(m: chess.Move) -> str:
return self._annotate_move_simple(board, m)
@ -336,8 +322,6 @@ class RandomEngine:
f"source=search depth={depth_used} time={elapsed:.2f}s nodes={getattr(self, '_nodes', 0)} candidates={len(scores)}",
f"best {board.san(top[0][0])} ({top[0][0].uci()}) score={best_cp:.1f} reasons=[{annotate(top[0][0])}]",
]
#if avoided_note:
# lines.append(avoided_note)
if len(top) > 1:
lines.append("alternatives:")
for mv, sc in top[1:]:
@ -408,6 +392,174 @@ class RandomEngine:
return best_move, "\n".join(lines)
def evaluate_proposed_move_with_suggestion(
self,
state: "chess.Board | str",
proposed_move: "chess.Move | str",
time_budget_sec: Optional[float] = None,
) -> Tuple[Optional[float], str, Optional[chess.Move], str]:
"""Evaluate a proposed move on a given position and also propose the engine's move.
Inputs:
- state: either a chess.Board or a FEN string representing the current position.
- proposed_move: a chess.Move object or a string (UCI preferred, SAN fallback).
- time_budget_sec: optional time limit for each analysis pass.
Returns a tuple: (proposed_score_cp, proposed_explanation, engine_move, engine_explanation)
- proposed_score_cp: search score (centipawns, mover POV) for the proposed move, or None if illegal/unavailable.
- proposed_explanation: a detailed, human-readable breakdown of where the score came from.
- engine_move: the move our engine would play from this position.
- engine_explanation: the usual explanation for the engine's chosen move.
"""
# Normalize board
if isinstance(state, chess.Board):
board = state.copy(stack=False)
else:
try:
board = chess.Board(state)
except Exception:
# If FEN parsing fails, assume start position
board = chess.Board()
# Parse proposed move
def _parse_move(b: chess.Board, mv: "chess.Move | str") -> Optional[chess.Move]:
if isinstance(mv, chess.Move):
return mv if mv in b.legal_moves else None
# Try UCI first
try:
m = chess.Move.from_uci(str(mv).strip())
if m in b.legal_moves:
return m
except Exception:
pass
# Try SAN
try:
m = b.parse_san(str(mv).strip())
if m in b.legal_moves:
return m
except Exception:
pass
return None
pmove = _parse_move(board, proposed_move)
# Evaluate proposed move (ensure we set a deadline specifically for this pass)
start = time.time()
budget = time_budget_sec if time_budget_sec is not None else min(1.5, self.max_time_sec)
self._deadline = start + max(0.01, budget)
self._nodes = 0
proposed_score: Optional[float] = None
if pmove is None:
# Still compute our own suggestion below
proposed_expl = (
"source=search\nillegal_or_unavailable_move=true\n"
"note=The proposed move is not legal in the provided position."
)
else:
# Compute search score for the specific move
# Use full alpha-beta for consistency with choose_move_with_explanation
# Depth is limited by deadline
depth_used = 0
best_score_for_move: Optional[float] = None
# A tiny iterative deepening loop focused on the single move
for d in range(1, self.depth + 1):
if time.time() >= self._deadline:
break
depth_used = d
board.push(pmove)
try:
sc = -self._alphabeta(board, d - 1, -float("inf"), float("inf"), start)
finally:
board.pop()
best_score_for_move = sc
proposed_score = best_score_for_move
# Build a detailed breakdown similar to choose_move_with_explanation
def _move_breakdown(b: chess.Board, m: chess.Move, search_score: float) -> str:
# SEE and risk
try:
see_val = int(self._see_value(b, m))
except Exception:
see_val = 0
risk_total = self._risk_score(b, m)
try:
risk_qtrap = self._queen_trap_risk(b, m)
except Exception:
risk_qtrap = 0
risk_bxf = 600 if (self._is_early_game(b) and self._is_bishop_sac_on_f2f7(b, m)) else 0
# Static eval components pre/post (mover perspective)
pre_white_score, pre_comp = self._evaluate_components(b)
pre_stm = pre_white_score if b.turn == chess.WHITE else -pre_white_score
b.push(m)
try:
post_white_score, post_comp = self._evaluate_components(b)
post_stm = - (post_white_score if b.turn == chess.WHITE else -post_white_score)
finally:
b.pop()
tactical_delta = search_score - post_stm
def fmt_comps(prefix: str, comp: dict, white_score_val: float, stm_val: float) -> list[str]:
parts = []
parts.append(f"{prefix}: stm_eval={stm_val:.1f} (from white_score={white_score_val:.1f} {'as-is' if (prefix=='pre') else 'flipped to mover'})")
parts.append(" components (white-centric):")
parts.append(f" material={comp['material']}")
parts.append(f" doubled_pawns_term={comp['doubled_pawns_term']}")
parts.append(f" mobility_term={comp['mobility_term']}")
parts.append(f" mobility_white={comp['mob_w']} mobility_black={comp['mob_b']}")
parts.append(f" center_score={comp['center_score']}")
parts.append(f" rook_file_bonus={comp['rook_file_bonus']}")
parts.append(f" king_safety={comp['safety']}")
parts.append(f" queen_raid_penalty={comp['queen_raid_pen']}")
parts.append(f" piece_square_table={comp['pst']}")
parts.append(f" hanging_pieces_term={comp['hanging_pieces_term']}")
return parts
pre_lines = fmt_comps("pre", pre_comp, pre_white_score, pre_stm)
post_lines = fmt_comps("post", post_comp, post_white_score, post_stm)
elapsed = time.time() - start
nodes = getattr(self, "_nodes", 0)
mv_san = None
try:
mv_san = b.san(m)
except Exception:
pass
ann = self._annotate_move_simple(b, m)
lines = [
f"source=search depth={depth_used} time={elapsed:.2f}s nodes={nodes}",
f"move {mv_san if mv_san else ''} ({m.uci()}) score={search_score:.1f} reasons=[{ann}]",
"details:",
f" see={see_val}",
f" risk_total={risk_total}",
f" risk_queen_trap={risk_qtrap}",
f" risk_bishop_sac_f2f7={risk_bxf}",
f" pre_static_eval: {pre_stm:.1f}",
f" post_static_eval: {post_stm:.1f}",
f" search_score: {search_score:.1f}",
f" tactical_delta: {tactical_delta:+.1f}",
]
lines.extend(pre_lines)
lines.extend(post_lines)
return "\n".join(lines)
if proposed_score is None:
proposed_expl = (
"source=search\nunable_to_compute_score=true\n"
"note=Time limit reached before evaluating the proposed move."
)
else:
proposed_expl = _move_breakdown(board, pmove, proposed_score)
# Now get the engine's own suggestion (separate pass to keep API simple)
eng_budget = time_budget_sec if time_budget_sec is not None else min(1.5, self.max_time_sec)
move_suggestion, engine_expl = self.choose_move_with_explanation(board, time_budget_sec=eng_budget)
return proposed_score, proposed_expl, move_suggestion, engine_expl
def _analyze_root(self, board: chess.Board, depth: int, start: float) -> list[Tuple[chess.Move, float]]:
alpha = -float("inf")
beta = float("inf")
@ -1184,376 +1336,3 @@ class RandomEngine:
finally:
board.pop()
return risk
# --- Blunder veto helpers ---
def _looks_blunderish(self, board: chess.Board, move: chess.Move) -> bool:
"""Lightweight tactical sanity checks to avoid common blunders.
Heuristics:
- Use a small DB of logged blunders
- Negative SEE on the move beyond a threshold (typically losing material)
- Large immediate static-eval drop
- Opponent has mate-in-1 or many forcing checks
- Opponent has profitable captures right away
- Destination under-defended and opponent can capture for non-negative SEE
- Tiny depth-2 probe is very favorable for the opponent
- Non-forced king moves in middlegame
"""
# Known logged blunder?
try:
if self._is_logged_blunder(board, move):
return True
except Exception:
pass
# Strongly negative SEE on our own move
try:
see = int(self._see_value(board, move))
except Exception:
see = 0
is_cap = board.is_capture(move)
if see <= -120:
return True
# Pawns: non-capture push into under-defended square is often a blunder even at milder SEE
pc0 = board.piece_at(move.from_square)
if pc0 and pc0.piece_type == chess.PAWN and not board.is_capture(move) and not move.promotion:
if see <= -60:
return True
from_file = chess.square_file(move.from_square)
if from_file in (0, 7):
heavy_pieces = sum(1 for p in board.piece_map().values() if p.piece_type in (chess.QUEEN, chess.ROOK))
if self._is_early_game(board) or heavy_pieces >= 2:
if see < 0:
return True
# Pre-move static eval (from mover perspective)
pre_eval = self._evaluate(board)
board.push(move)
try:
# Post-move eval (flip sign back to mover perspective)
post_eval_for_opp = self._evaluate(board)
post_eval_for_us = -post_eval_for_opp
# For captures with non-negative SEE, require a much larger immediate static drop to flag as a blunder
drop_thresh = 350 if (is_cap and see >= 0) else 220
if post_eval_for_us < pre_eval - drop_thresh:
return True
# Opponent mate-in-1?
for opp in board.legal_moves:
board.push(opp)
try:
if board.is_checkmate():
return True
finally:
board.pop()
# Forcing checks right away
check_count = 0
heavy_check = False
my_color = not board.turn
for opp in board.legal_moves:
try:
if board.gives_check(opp):
check_count += 1
pc = board.piece_at(opp.from_square)
if pc and pc.piece_type in (chess.QUEEN, chess.ROOK, chess.BISHOP):
heavy_check = True
if board.is_capture(opp) or opp.promotion:
heavy_check = True
except Exception:
pass
# For captures that pass SEE, be less sensitive to raw check counts
if (not (is_cap and see >= 0) and (check_count >= 2 or (check_count >= 1 and heavy_check))) or ((is_cap and see >= 0) and check_count >= 3):
return True
# King exits limited after check presence
ksq = board.king(my_color)
if ksq is not None and check_count >= 1:
king_exits = 0
for m in board.legal_moves:
if m.from_square == ksq and not board.is_capture(m):
king_exits += 1
if king_exits >= 2:
break
if king_exits <= 1:
return True
# Under-defended destination
moved_piece = board.piece_at(move.to_square)
if moved_piece:
# If this was a capture with non-negative SEE, trust SEE over crude defenders/attackers
if not (is_cap and see >= 0):
attackers_op = len(board.attackers(not my_color, move.to_square))
defenders_me = len(board.attackers(my_color, move.to_square))
if attackers_op >= max(1, defenders_me):
for opp in board.legal_moves:
if opp.to_square == move.to_square and board.is_capture(opp):
try:
opp_see2 = int(self._see_value(board, opp))
except Exception:
opp_see2 = 0
if opp_see2 >= 0:
return True
# Greedy corner-rook takes by minors are often traps (similar to queen corner traps)
if is_cap:
moved_by = board.piece_at(move.to_square)
from_pc = board.piece_at(move.from_square) if moved_by is None else moved_by # after push, piece on to_square
victim_pre = None
# Reconstruct victim type by undoing and checking
board.pop()
try:
victim_pre = board.piece_at(move.to_square)
finally:
board.push(move)
if from_pc and from_pc.piece_type in (chess.KNIGHT, chess.BISHOP) and victim_pre and victim_pre.piece_type == chess.ROOK and move.to_square in {chess.A8, chess.H8, chess.A1, chess.H1}:
# If destination is under-defended OR the minor has very limited safe exits, treat as blunderish
attackers = len(board.attackers(not my_color, move.to_square))
defenders = len(board.attackers(my_color, move.to_square))
# Count safe exits for the minor from the corner square
safe_exits = 0
for esc in board.legal_moves:
if esc.from_square == move.to_square:
# Avoid landing on attacked squares
if not board.is_attacked_by(not my_color, esc.to_square):
safe_exits += 1
if safe_exits >= 3:
break
if attackers >= max(1, defenders) or safe_exits <= 1:
return True
# Unsafe promotions to queen: if the new queen square is under-defended, treat as blunderish
if move.promotion == chess.QUEEN:
qsq = move.to_square
attackers = len(board.attackers(not my_color, qsq))
defenders = len(board.attackers(my_color, qsq))
if attackers >= max(1, defenders):
return True
# Rook-pawn loosening immediately increasing own hanging penalties is a red flag
mover_from_file = chess.square_file(move.from_square)
if moved_piece and moved_piece.piece_type == chess.PAWN and mover_from_file in (0, 7) and not board.is_capture(move) and not move.promotion:
_, pre_comp = self._evaluate_components(board) # board here is after push
pre_hanging_term = pre_comp.get("hanging_pieces_term", 0)
# Undo move to get baseline, then re-push for state safety
board.pop()
try:
_, base_comp = self._evaluate_components(board)
base_hanging_term = base_comp.get("hanging_pieces_term", 0)
finally:
board.push(move)
# If our side's hanging penalty worsened notably, treat as blunder-ish
# Remember: components are white-centric; flip based on mover color
delta_hang_white = pre_hanging_term - base_hanging_term
worsen = (delta_hang_white < -80) if my_color == chess.WHITE else (delta_hang_white > 80)
if worsen:
return True
# Tiny probe depth-2 for opponent side
try:
old_deadline = getattr(self, "_deadline", None)
self._deadline = time.time() + 0.03
probe = self._alphabeta(board, 2, -float('inf'), float('inf'), time.time())
if old_deadline is not None:
self._deadline = old_deadline
# For safe captures (SEE >= 0), require a much stronger opponent probe to flag as blunder
# Only relax the tiny-probe threshold for queen captures that already passed SEE
relax = is_cap and see >= 0 and board.piece_at(move.from_square) and board.piece_at(move.from_square).piece_type == chess.QUEEN
threshold = 450 if relax else 250
if probe >= threshold:
return True
except Exception:
try:
if old_deadline is not None:
self._deadline = old_deadline
except Exception:
pass
finally:
board.pop()
# Non-forced king move in middlegame
pc0 = board.piece_at(move.from_square)
if pc0 and pc0.piece_type == chess.KING and not board.is_castling(move):
# If we're currently in check, allow king moves (forced). Otherwise, scrutinize.
if not board.is_check():
# Penalize king moves that keep the king in the center early or that stay on the same file/rank while heavy pieces remain.
heavy_pieces = sum(1 for p in board.piece_map().values() if p.piece_type in (chess.QUEEN, chess.ROOK))
if self._is_early_game(board) or heavy_pieces >= 2:
# Compute a crude safety trend: if post static eval (mover perspective) does not improve and king remains central/edge-trapped, veto.
pre_eval = self._evaluate(board)
board.push(move)
try:
post_eval_for_opp = self._evaluate(board)
post_eval_for_us = -post_eval_for_opp
to_sq = move.to_square
file_idx = chess.square_file(to_sq)
rank_idx = chess.square_rank(to_sq)
centralish = (file_idx in (3, 4)) and (rank_idx in (2, 3, 4, 5))
stays_corner = (to_sq in {chess.A1, chess.H1, chess.A8, chess.H8})
# If not improving eval and moving into/remaining in danger zones, veto.
if post_eval_for_us <= pre_eval and (not centralish or stays_corner):
return True
finally:
board.pop()
# Generic: non-forced king shuffle is discouraged
return True
return False
def _pick_safer_alternative(self, board: chess.Board, avoid: Optional[chess.Move] = None) -> Optional[chess.Move]:
"""Pick a safer alternative move using ordering heuristics, avoiding a specific move if provided.
Prefers non-logged, non-blunderish moves; falls back to the least-risk option.
"""
moves = [m for m in self._ordered_moves(board) if avoid is None or m != avoid]
if not moves:
return None
def is_non_forced_king_move(mv: chess.Move) -> bool:
pc = board.piece_at(mv.from_square)
if not pc or pc.piece_type != chess.KING:
return False
if board.is_castling(mv):
return False
# Non-check, non-capture king moves are considered non-forced
if board.is_capture(mv):
return False
try:
if board.gives_check(mv):
return False
except Exception:
pass
# Avoid passive king shuffles when heavy pieces remain
heavy_pieces = sum(1 for p in board.piece_map().values() if p.piece_type in (chess.QUEEN, chess.ROOK))
return heavy_pieces >= 2 or self._is_early_game(board)
def is_plain_pawn_push(mv: chess.Move) -> bool:
pc = board.piece_at(mv.from_square)
if not pc or pc.piece_type != chess.PAWN:
return False
if board.is_capture(mv) or mv.promotion:
return False
try:
if board.gives_check(mv):
return False
except Exception:
pass
return True
# First pass: strictly avoid logged blunders and blunderish moves; also skip passive king shuffles if possible
# Prefer SEE-sound captures with the lowest risk (avoid unsafe queen promotions)
safe_caps: list[tuple[int, chess.Move]] = []
for m in moves:
if board.is_capture(m):
try:
see_m = int(self._see_value(board, m))
except Exception:
see_m = 0
if see_m >= 0:
try:
if self._is_logged_blunder(board, m):
continue
except Exception:
pass
try:
r = self._risk_score(board, m)
except Exception:
r = 9999
# Small tie-breaker: prefer non-queen promotions over queen when risks tie
if m.promotion == chess.QUEEN:
r += 50
safe_caps.append((r, m))
if safe_caps:
safe_caps.sort(key=lambda t: t[0])
return safe_caps[0][1]
# Otherwise look for solid non-captures following safety heuristics
for m in moves:
try:
if self._is_logged_blunder(board, m):
continue
except Exception:
pass
if is_non_forced_king_move(m):
continue
# If many minors are undeveloped, prefer non-pawn moves first
mover = board.piece_at(m.from_square)
if mover and self._most_minors_undeveloped(board, mover.color) and is_plain_pawn_push(m):
continue
# Prefer captures or developing moves over rook-pawn pushes
if mover and mover.piece_type == chess.PAWN and is_plain_pawn_push(m):
from_file = chess.square_file(m.from_square)
if from_file in (0, 7):
continue
if not self._looks_blunderish(board, m):
return m
# Second pass: choose least-risk non-logged, non-blunderish move; allow passive king last
scored: list[tuple[int, chess.Move]] = []
for m in moves:
try:
if self._is_logged_blunder(board, m):
continue
except Exception:
pass
# Skip blunderish moves outright, but allow passive king moves as last resort (still add risk)
try:
bl = self._looks_blunderish(board, m)
except Exception:
bl = False
if bl and not is_non_forced_king_move(m):
continue
try:
r = self._risk_score(board, m)
except Exception:
r = 9999
# Slightly inflate risk for passive king moves to avoid endless shuffling when heavy pieces remain
if is_non_forced_king_move(m):
r += 250
# Inflate risk for plain pawn pushes if minors are still undeveloped
mover = board.piece_at(m.from_square)
if mover and self._most_minors_undeveloped(board, mover.color) and is_plain_pawn_push(m):
r += 180
# Extra risk for rook-pawn pushes in general
if mover and mover.piece_type == chess.PAWN and is_plain_pawn_push(m):
if chess.square_file(m.from_square) in (0, 7):
r += 120
# And extra for SEE-negative plain pawn advances
try:
see_m = int(self._see_value(board, m))
except Exception:
see_m = 0
if see_m < 0:
r += 220
scored.append((r, m))
if scored:
scored.sort(key=lambda t: t[0])
return scored[0][1]
# Last resort: still avoid logged blunders if at all possible
non_logged = []
for m in moves:
try:
if self._is_logged_blunder(board, m):
continue
except Exception:
pass
non_logged.append(m)
if not non_logged and avoid is not None:
# Better to take the previously avoided (but not logged) move than a known logged blunder
try:
if not self._is_logged_blunder(board, avoid):
return avoid
except Exception:
return avoid
target_pool = non_logged if non_logged else moves
try:
target_pool.sort(key=lambda m: self._risk_score(board, m))
except Exception:
pass
return target_pool[0]
def _is_logged_blunder(self, board: chess.Board, move: chess.Move) -> bool:
fen = board.fen()
bad = self._logged_blunders.get(fen)
return bool(bad and move.uci() in bad)