X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsearch.cpp;h=3b25b6533db24730049f20f156ebffa6b6e82bd3;hp=09d272e56983361fe813032c0051adf011169aac;hb=e8e5b9f537a4b8b062b48f584f184f254385b276;hpb=55bd27b8f08a151128d7065fa2819aa3e9605299 diff --git a/src/search.cpp b/src/search.cpp index 09d272e5..3b25b653 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -42,6 +42,7 @@ namespace Search { LimitsType Limits; std::vector RootMoves; Position RootPosition; + Color RootColor; Time::point SearchTime; StateStackPtr SetupStates; } @@ -89,9 +90,8 @@ namespace { size_t MultiPV, UCIMultiPV, PVIdx; TimeManager TimeMgr; int BestMoveChanges; - int SkillLevel; - bool SkillLevelEnabled, Chess960; - Value DrawValue[2]; + bool Chess960; + Value DrawValue[COLOR_NB]; History H; template @@ -106,10 +106,24 @@ namespace { Value value_to_tt(Value v, int ply); Value value_from_tt(Value v, int ply); bool connected_threat(const Position& pos, Move m, Move threat); - Value refine_eval(const TTEntry* tte, Value ttValue, Value defaultEval); - Move do_skill_level(); string uci_pv(const Position& pos, int depth, Value alpha, Value beta); + struct Skill { + Skill(int l) : level(l), best(MOVE_NONE) {} + ~Skill() { + if (enabled()) // Swap best PV line with the sub-optimal one + std::swap(RootMoves[0], *std::find(RootMoves.begin(), + RootMoves.end(), best ? best : pick_move())); + } + + bool enabled() const { return level < 20; } + bool time_to_pick(int depth) const { return depth == 1 + level; } + Move pick_move(); + + int level; + Move best; + }; + } // namespace @@ -174,7 +188,7 @@ void Search::think() { Position& pos = RootPosition; Chess960 = pos.is_chess960(); - Eval::RootColor = pos.side_to_move(); + RootColor = pos.side_to_move(); TimeMgr.init(Limits, pos.startpos_ply_counter(), pos.side_to_move()); TT.new_search(); H.clear(); @@ -192,8 +206,8 @@ void Search::think() { { int cf = Options["Contempt Factor"] * PawnValueMg / 100; // In centipawns cf = cf * MaterialTable::game_phase(pos) / PHASE_MIDGAME; // Scale down with phase - DrawValue[ Eval::RootColor] = VALUE_DRAW - Value(cf); - DrawValue[~Eval::RootColor] = VALUE_DRAW + Value(cf); + DrawValue[ RootColor] = VALUE_DRAW - Value(cf); + DrawValue[~RootColor] = VALUE_DRAW + Value(cf); } else DrawValue[WHITE] = DrawValue[BLACK] = VALUE_DRAW; @@ -209,14 +223,6 @@ void Search::think() { } } - UCIMultiPV = Options["MultiPV"]; - SkillLevel = Options["Skill Level"]; - - // Do we have to play with skill handicap? In this case enable MultiPV that - // we will use behind the scenes to retrieve a set of possible moves. - SkillLevelEnabled = (SkillLevel < 20); - MultiPV = (SkillLevelEnabled ? std::max(UCIMultiPV, (size_t)4) : UCIMultiPV); - if (Options["Use Search Log"]) { Log log(Options["Search Log Filename"]); @@ -287,15 +293,21 @@ namespace { int depth, prevBestMoveChanges; Value bestValue, alpha, beta, delta; bool bestMoveNeverChanged = true; - Move skillBest = MOVE_NONE; memset(ss, 0, 4 * sizeof(Stack)); depth = BestMoveChanges = 0; bestValue = delta = -VALUE_INFINITE; ss->currentMove = MOVE_NULL; // Hack to skip update gains + UCIMultiPV = Options["MultiPV"]; + Skill skill(Options["Skill Level"]); + + // Do we have to play with skill handicap? In this case enable MultiPV that + // we will use behind the scenes to retrieve a set of possible moves. + MultiPV = skill.enabled() ? std::max(UCIMultiPV, (size_t)4) : UCIMultiPV; + // Iterative deepening loop until requested to stop or target depth reached - while (!Signals.stop && ++depth <= MAX_PLY && (!Limits.depth || depth <= Limits.depth)) + while (++depth <= MAX_PLY && !Signals.stop && (!Limits.depth || depth <= Limits.depth)) { // Save last iteration's scores before first PV line is searched and all // the move scores but the (new) PV are set to -VALUE_INFINITE. @@ -337,37 +349,37 @@ namespace { // the already searched PV lines are preserved. sort(RootMoves.begin() + PVIdx, RootMoves.end()); - // In case we have found an exact score and we are going to leave - // the fail high/low loop then reorder the PV moves, otherwise - // leave the last PV move in its position so to be searched again. - // Of course this is needed only in MultiPV search. - if (PVIdx && bestValue > alpha && bestValue < beta) - sort(RootMoves.begin(), RootMoves.begin() + PVIdx); - // Write PV back to transposition table in case the relevant // entries have been overwritten during the search. for (size_t i = 0; i <= PVIdx; i++) RootMoves[i].insert_pv_in_tt(pos); - // If search has been stopped exit the aspiration window loop. - // Sorting and writing PV back to TT is safe becuase RootMoves - // is still valid, although refers to previous iteration. + // If search has been stopped return immediately. Sorting and + // writing PV back to TT is safe becuase RootMoves is still + // valid, although refers to previous iteration. if (Signals.stop) + return; + + // In case of failing high/low increase aspiration window and + // research, otherwise exit the loop. + if (bestValue > alpha && bestValue < beta) break; - // Send full PV info to GUI if we are going to leave the loop or - // if we have a fail high/low and we are deep in the search. - if ((bestValue > alpha && bestValue < beta) || Time::now() - SearchTime > 2000) + // Give some update (without cluttering the UI) before to research + if (Time::now() - SearchTime > 3000) sync_cout << uci_pv(pos, depth, alpha, beta) << sync_endl; - // In case of failing high/low increase aspiration window and - // research, otherwise exit the fail high/low loop. - if (bestValue >= beta) + if (abs(bestValue) >= VALUE_KNOWN_WIN) + { + alpha = -VALUE_INFINITE; + beta = VALUE_INFINITE; + } + else if (bestValue >= beta) { beta += delta; delta += delta / 2; } - else if (bestValue <= alpha) + else { Signals.failedLowAtRoot = true; Signals.stopOnPonderhit = false; @@ -375,25 +387,20 @@ namespace { alpha -= delta; delta += delta / 2; } - else - break; - - // Search with full window in case we have a win/mate score - if (abs(bestValue) >= VALUE_KNOWN_WIN) - { - alpha = -VALUE_INFINITE; - beta = VALUE_INFINITE; - } assert(alpha >= -VALUE_INFINITE && beta <= VALUE_INFINITE); } + + // Sort the PV lines searched so far and update the GUI + sort(RootMoves.begin(), RootMoves.begin() + PVIdx); + sync_cout << uci_pv(pos, depth, alpha, beta) << sync_endl; } - // Skills: Do we need to pick now the best move ? - if (SkillLevelEnabled && depth == 1 + SkillLevel) - skillBest = do_skill_level(); + // Do we need to pick now the sub-optimal best move ? + if (skill.enabled() && skill.time_to_pick(depth)) + skill.pick_move(); - if (!Signals.stop && Options["Use Search Log"]) + if (Options["Use Search Log"]) { Log log(Options["Search Log Filename"]); log << pretty_pv(pos, depth, bestValue, Time::now() - SearchTime, &RootMoves[0].pv[0]) @@ -405,7 +412,7 @@ namespace { bestMoveNeverChanged = false; // Do we have time for the next iteration? Can we stop searching now? - if (!Signals.stop && !Signals.stopOnPonderhit && Limits.use_time_management()) + if (Limits.use_time_management() && !Signals.stopOnPonderhit) { bool stop = false; // Local variable, not the volatile Signals.stop @@ -447,15 +454,6 @@ namespace { } } } - - // When using skills swap best PV line with the sub-optimal one - if (SkillLevelEnabled) - { - if (skillBest == MOVE_NONE) // Still unassigned ? - skillBest = do_skill_level(); - - std::swap(RootMoves[0], *std::find(RootMoves.begin(), RootMoves.end(), skillBest)); - } } @@ -485,7 +483,7 @@ namespace { Move ttMove, move, excludedMove, bestMove, threatMove; Depth ext, newDepth; Value bestValue, value, ttValue; - Value refinedValue, nullValue, futilityValue; + Value eval, nullValue, futilityValue; bool inCheck, givesCheck, pvMove, singularExtensionNode; bool captureOrPromotion, dangerous, doFullDepthSearch; int moveCount, playedMoveCount; @@ -575,40 +573,45 @@ namespace { // Step 5. Evaluate the position statically and update parent's gain statistics if (inCheck) - ss->eval = ss->evalMargin = refinedValue = VALUE_NONE; + ss->staticEval = ss->evalMargin = eval = VALUE_NONE; else if (tte) { assert(tte->static_value() != VALUE_NONE); + assert(ttValue != VALUE_NONE || tte->type() == BOUND_NONE); - ss->eval = tte->static_value(); + ss->staticEval = eval = tte->static_value(); ss->evalMargin = tte->static_value_margin(); - refinedValue = refine_eval(tte, ttValue, ss->eval); + + // Can ttValue be used as a better position evaluation? + if ( ((tte->type() & BOUND_LOWER) && ttValue > eval) + || ((tte->type() & BOUND_UPPER) && ttValue < eval)) + eval = ttValue; } else { - refinedValue = ss->eval = evaluate(pos, ss->evalMargin); + eval = ss->staticEval = evaluate(pos, ss->evalMargin); TT.store(posKey, VALUE_NONE, BOUND_NONE, DEPTH_NONE, MOVE_NONE, - ss->eval, ss->evalMargin); + ss->staticEval, ss->evalMargin); } // Update gain for the parent non-capture move given the static position // evaluation before and after the move. - if ( (move = (ss-1)->currentMove) != MOVE_NULL - && (ss-1)->eval != VALUE_NONE - && ss->eval != VALUE_NONE + if ( (move = (ss-1)->currentMove) != MOVE_NULL + && (ss-1)->staticEval != VALUE_NONE + && ss->staticEval != VALUE_NONE && !pos.captured_piece_type() && type_of(move) == NORMAL) { Square to = to_sq(move); - H.update_gain(pos.piece_on(to), to, -(ss-1)->eval - ss->eval); + H.update_gain(pos.piece_on(to), to, -(ss-1)->staticEval - ss->staticEval); } // Step 6. Razoring (is omitted in PV nodes) if ( !PvNode && depth < 4 * ONE_PLY && !inCheck - && refinedValue + razor_margin(depth) < beta + && eval + razor_margin(depth) < beta && ttMove == MOVE_NONE && abs(beta) < VALUE_MATE_IN_MAX_PLY && !pos.pawn_on_7th(pos.side_to_move())) @@ -628,17 +631,17 @@ namespace { && !ss->skipNullMove && depth < 4 * ONE_PLY && !inCheck - && refinedValue - FutilityMargins[depth][0] >= beta + && eval - FutilityMargins[depth][0] >= beta && abs(beta) < VALUE_MATE_IN_MAX_PLY && pos.non_pawn_material(pos.side_to_move())) - return refinedValue - FutilityMargins[depth][0]; + return eval - FutilityMargins[depth][0]; // Step 8. Null move search with verification search (is omitted in PV nodes) if ( !PvNode && !ss->skipNullMove && depth > ONE_PLY && !inCheck - && refinedValue >= beta + && eval >= beta && abs(beta) < VALUE_MATE_IN_MAX_PLY && pos.non_pawn_material(pos.side_to_move())) { @@ -648,7 +651,7 @@ namespace { Depth R = 3 * ONE_PLY + depth / 4; // Null move dynamic reduction based on value - if (refinedValue - PawnValueMg > beta) + if (eval - PawnValueMg > beta) R += ONE_PLY; pos.do_null_move(st); @@ -729,7 +732,7 @@ namespace { // Step 10. Internal iterative deepening if ( depth >= (PvNode ? 5 * ONE_PLY : 8 * ONE_PLY) && ttMove == MOVE_NONE - && (PvNode || (!inCheck && ss->eval + Value(256) >= beta))) + && (PvNode || (!inCheck && ss->staticEval + Value(256) >= beta))) { Depth d = (PvNode ? depth - 2 * ONE_PLY : depth / 2); @@ -801,7 +804,7 @@ split_point_start: // At split points actual search starts from here && type_of(pos.piece_on(to_sq(move))) != PAWN && type_of(move) == NORMAL && ( pos.non_pawn_material(WHITE) + pos.non_pawn_material(BLACK) - - PieceValue[Mg][pos.piece_on(to_sq(move))] == VALUE_ZERO)); + - PieceValue[MG][pos.piece_on(to_sq(move))] == VALUE_ZERO)); // Step 12. Extend checks and, in PV nodes, also dangerous moves if (PvNode && dangerous) @@ -816,8 +819,8 @@ split_point_start: // At split points actual search starts from here // on all the other moves but the ttMove, if result is lower than ttValue minus // a margin then we extend ttMove. if ( singularExtensionNode - && !ext && move == ttMove + && !ext && pos.pl_move_is_legal(move, ci.pinned) && abs(ttValue) < VALUE_KNOWN_WIN) { @@ -861,7 +864,7 @@ split_point_start: // At split points actual search starts from here // We illogically ignore reduction condition depth >= 3*ONE_PLY for predicted depth, // but fixing this made program slightly weaker. Depth predictedDepth = newDepth - reduction(depth, moveCount); - futilityValue = ss->eval + ss->evalMargin + futility_margin(predictedDepth, moveCount) + futilityValue = ss->staticEval + ss->evalMargin + futility_margin(predictedDepth, moveCount) + H.gain(pos.piece_moved(move), to_sq(move)); if (futilityValue < beta) @@ -1037,7 +1040,7 @@ split_point_start: // At split points actual search starts from here if (bestValue >= beta) // Failed high { TT.store(posKey, value_to_tt(bestValue, ss->ply), BOUND_LOWER, depth, - bestMove, ss->eval, ss->evalMargin); + bestMove, ss->staticEval, ss->evalMargin); if (!pos.is_capture_or_promotion(bestMove) && !inCheck) { @@ -1062,7 +1065,7 @@ split_point_start: // At split points actual search starts from here else // Failed low or PV search TT.store(posKey, value_to_tt(bestValue, ss->ply), PvNode && bestMove != MOVE_NONE ? BOUND_EXACT : BOUND_UPPER, - depth, bestMove, ss->eval, ss->evalMargin); + depth, bestMove, ss->staticEval, ss->evalMargin); assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1126,7 +1129,7 @@ split_point_start: // At split points actual search starts from here // Evaluate the position statically if (inCheck) { - ss->eval = ss->evalMargin = VALUE_NONE; + ss->staticEval = ss->evalMargin = VALUE_NONE; bestValue = futilityBase = -VALUE_INFINITE; enoughMaterial = false; } @@ -1136,18 +1139,18 @@ split_point_start: // At split points actual search starts from here { assert(tte->static_value() != VALUE_NONE); - ss->eval = bestValue = tte->static_value(); + ss->staticEval = bestValue = tte->static_value(); ss->evalMargin = tte->static_value_margin(); } else - ss->eval = bestValue = evaluate(pos, ss->evalMargin); + ss->staticEval = bestValue = evaluate(pos, ss->evalMargin); // Stand pat. Return immediately if static value is at least beta if (bestValue >= beta) { if (!tte) TT.store(pos.key(), value_to_tt(bestValue, ss->ply), BOUND_LOWER, - DEPTH_NONE, MOVE_NONE, ss->eval, ss->evalMargin); + DEPTH_NONE, MOVE_NONE, ss->staticEval, ss->evalMargin); return bestValue; } @@ -1155,7 +1158,7 @@ split_point_start: // At split points actual search starts from here if (PvNode && bestValue > alpha) alpha = bestValue; - futilityBase = ss->eval + ss->evalMargin + Value(128); + futilityBase = ss->staticEval + ss->evalMargin + Value(128); enoughMaterial = pos.non_pawn_material(pos.side_to_move()) > RookValueMg; } @@ -1183,7 +1186,7 @@ split_point_start: // At split points actual search starts from here && !pos.is_passed_pawn_push(move)) { futilityValue = futilityBase - + PieceValue[Eg][pos.piece_on(to_sq(move))] + + PieceValue[EG][pos.piece_on(to_sq(move))] + (type_of(move) == ENPASSANT ? PawnValueEg : VALUE_ZERO); if (futilityValue < beta) @@ -1222,7 +1225,7 @@ split_point_start: // At split points actual search starts from here && givesCheck && move != ttMove && !pos.is_capture_or_promotion(move) - && ss->eval + PawnValueMg / 4 < beta + && ss->staticEval + PawnValueMg / 4 < beta && !check_is_dangerous(pos, move, futilityBase, beta)) continue; @@ -1254,7 +1257,7 @@ split_point_start: // At split points actual search starts from here else // Fail high { TT.store(posKey, value_to_tt(value, ss->ply), BOUND_LOWER, - ttDepth, move, ss->eval, ss->evalMargin); + ttDepth, move, ss->staticEval, ss->evalMargin); return value; } @@ -1269,7 +1272,7 @@ split_point_start: // At split points actual search starts from here TT.store(posKey, value_to_tt(bestValue, ss->ply), PvNode && bestMove != MOVE_NONE ? BOUND_EXACT : BOUND_UPPER, - ttDepth, bestMove, ss->eval, ss->evalMargin); + ttDepth, bestMove, ss->staticEval, ss->evalMargin); assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1314,7 +1317,7 @@ split_point_start: // At split points actual search starts from here while (b) { // Note that here we generate illegal "double move"! - if (futilityBase + PieceValue[Eg][pos.piece_on(pop_lsb(&b))] >= beta) + if (futilityBase + PieceValue[EG][pos.piece_on(pop_lsb(&b))] >= beta) return true; } @@ -1419,7 +1422,7 @@ split_point_start: // At split points actual search starts from here // Case 2: If the threatened piece has value less than or equal to the // value of the threatening piece, don't prune moves which defend it. if ( pos.is_capture(threat) - && ( PieceValue[Mg][pos.piece_on(tfrom)] >= PieceValue[Mg][pos.piece_on(tto)] + && ( PieceValue[MG][pos.piece_on(tfrom)] >= PieceValue[MG][pos.piece_on(tto)] || type_of(pos.piece_on(tfrom)) == KING) && pos.move_attacks_square(m, tto)) return true; @@ -1435,27 +1438,10 @@ split_point_start: // At split points actual search starts from here } - // refine_eval() returns the transposition table score if possible, otherwise - // falls back on static position evaluation. Note that we never return VALUE_NONE - // even if v == VALUE_NONE. - - Value refine_eval(const TTEntry* tte, Value v, Value defaultEval) { - - assert(tte); - assert(v != VALUE_NONE || !tte->type()); - - if ( ((tte->type() & BOUND_LOWER) && v >= defaultEval) - || ((tte->type() & BOUND_UPPER) && v < defaultEval)) - return v; - - return defaultEval; - } - - // When playing with strength handicap choose best move among the MultiPV set - // using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen. + // using a statistical rule dependent on 'level'. Idea by Heinz van Saanen. - Move do_skill_level() { + Move Skill::pick_move() { assert(MultiPV > 1); @@ -1468,9 +1454,9 @@ split_point_start: // At split points actual search starts from here // RootMoves are already sorted by score in descending order size_t size = std::min(MultiPV, RootMoves.size()); int variance = std::min(RootMoves[0].score - RootMoves[size - 1].score, PawnValueMg); - int weakness = 120 - 2 * SkillLevel; + int weakness = 120 - 2 * level; int max_s = -VALUE_INFINITE; - Move best = MOVE_NONE; + best = MOVE_NONE; // Choose best move. For each move score we add two terms both dependent on // weakness, one deterministic and bigger for weaker moves, and one random,