kate/BadFish
1
0
Fork 0
mirror of https://github.com/sockspls/badfish synced 2025-04-30 08:43:09 +00:00
Code Issues Releases Wiki Activity

Better document how MultiPV search works

Browse source 
No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
This commit is contained in:
Marco Costalba 2011-10-08 10:51:39 +01:00
parent a3bf09c5c9
commit fec623d68d
1 changed files with 44 additions and 40 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

84
src/search.cpp
View file

@ -159,7 +159,7 @@ namespace {
RootMoveList Rml; RootMoveList Rml;
// MultiPV mode // MultiPV mode
int MultiPV, UCIMultiPV, MultiPVIteration; int MultiPV, UCIMultiPV, MultiPVIdx;
// Time management variables // Time management variables
bool StopOnPonderhit, FirstRootMove, StopRequest, QuitRequest, AspirationFailLow; bool StopOnPonderhit, FirstRootMove, StopRequest, QuitRequest, AspirationFailLow;
@ -527,18 +527,17 @@ namespace {
// Iterative deepening loop until requested to stop or target depth reached // Iterative deepening loop until requested to stop or target depth reached
while (!StopRequest && ++depth <= PLY_MAX && (!Limits.maxDepth || depth <= Limits.maxDepth)) while (!StopRequest && ++depth <= PLY_MAX && (!Limits.maxDepth || depth <= Limits.maxDepth))
{ {
// Save last iteration's scores, this needs to be done now, because in // Save now last iteration's scores, before Rml moves are reordered
// the following MultiPV loop Rml moves could be reordered.
for (size_t i = 0; i < Rml.size(); i++) for (size_t i = 0; i < Rml.size(); i++)
Rml[i].prevScore = Rml[i].score; Rml[i].prevScore = Rml[i].score;
Rml.bestMoveChanges = 0; Rml.bestMoveChanges = 0;
// MultiPV iteration loop // MultiPV loop. We perform a full root search for each PV line
for (MultiPVIteration = 0; MultiPVIteration < Min(MultiPV, (int)Rml.size()); MultiPVIteration++) for (MultiPVIdx = 0; MultiPVIdx < Min(MultiPV, (int)Rml.size()); MultiPVIdx++)
{ {
// Calculate dynamic aspiration window based on previous iterations // Calculate dynamic aspiration window based on previous iterations
if (depth >= 5 && abs(Rml[MultiPVIteration].prevScore) < VALUE_KNOWN_WIN) if (depth >= 5 && abs(Rml[MultiPVIdx].prevScore) < VALUE_KNOWN_WIN)
{ {
int prevDelta1 = bestValues[depth - 1] - bestValues[depth - 2]; int prevDelta1 = bestValues[depth - 1] - bestValues[depth - 2];
int prevDelta2 = bestValues[depth - 2] - bestValues[depth - 3]; int prevDelta2 = bestValues[depth - 2] - bestValues[depth - 3];
@ -546,8 +545,8 @@ namespace {
aspirationDelta = Min(Max(abs(prevDelta1) + abs(prevDelta2) / 2, 16), 24); aspirationDelta = Min(Max(abs(prevDelta1) + abs(prevDelta2) / 2, 16), 24);
aspirationDelta = (aspirationDelta + 7) / 8 * 8; // Round to match grainSize aspirationDelta = (aspirationDelta + 7) / 8 * 8; // Round to match grainSize
alpha = Max(Rml[MultiPVIteration].prevScore - aspirationDelta, -VALUE_INFINITE); alpha = Max(Rml[MultiPVIdx].prevScore - aspirationDelta, -VALUE_INFINITE);
beta = Min(Rml[MultiPVIteration].prevScore + aspirationDelta, VALUE_INFINITE); beta = Min(Rml[MultiPVIdx].prevScore + aspirationDelta, VALUE_INFINITE);
} }
else else
{ {
@ -558,39 +557,44 @@ namespace {
// Start with a small aspiration window and, in case of fail high/low, // Start with a small aspiration window and, in case of fail high/low,
// research with bigger window until not failing high/low anymore. // research with bigger window until not failing high/low anymore.
do { do {
// Search starting from ss+1 to allow referencing (ss-1). This is // Search starts from ss+1 to allow referencing (ss-1). This is
// needed by update_gains() and ss copy when splitting at Root. // needed by update_gains() and ss copy when splitting at Root.
value = search<Root>(pos, ss+1, alpha, beta, depth * ONE_PLY); value = search<Root>(pos, ss+1, alpha, beta, depth * ONE_PLY);
// It is critical that sorting is done with a stable algorithm // Bring to front the best move. It is critical that sorting is
// because all the values but the first are usually set to // done with a stable algorithm because all the values but the first
// -VALUE_INFINITE and we want to keep the same order for all // and eventually the new best one are set to -VALUE_INFINITE and
// the moves but the new PV that goes to head. // we want to keep the same order for all the moves but the new
sort<RootMove>(Rml.begin() + MultiPVIteration, Rml.end()); // PV that goes to the front. Note that in case of MultiPV search
// the already searched PV lines are preserved.
sort<RootMove>(Rml.begin() + MultiPVIdx, Rml.end());
// In case we have found an exact score reorder the PV moves // In case we have found an exact score and we are going to leave
// before leaving the fail high/low loop, otherwise leave the // the fail high/low loop then reorder the PV moves, otherwise
// last PV move in its position so to be searched again. // leave the last PV move in its position so to be searched again.
if (value > alpha && value < beta) // Of course this is needed only in MultiPV search.
sort<RootMove>(Rml.begin(), Rml.begin() + MultiPVIteration); if (MultiPVIdx && value > alpha && value < beta)
sort<RootMove>(Rml.begin(), Rml.begin() + MultiPVIdx);
// Write PV back to transposition table in case the relevant entries // Write PV back to transposition table in case the relevant entries
// have been overwritten during the search. // have been overwritten during the search.
for (int i = 0; i <= MultiPVIteration; i++) for (int i = 0; i <= MultiPVIdx; i++)
Rml[i].insert_pv_in_tt(pos); Rml[i].insert_pv_in_tt(pos);
// Value cannot be trusted. Break out immediately! // If search has been stopped exit the aspiration window loop,
// note that sorting and writing PV back to TT is safe becuase
// Rml is still valid, although refers to the previous iteration.
if (StopRequest) if (StopRequest)
break; break;
// Send full PV info to GUI if we are going to leave the loop or // 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. Note // if we have a fail high/low and we are deep in the search. UCI
// that UCI protol requires to send all the PV lines also if are // protocol requires to send all the PV lines also if are still
// still to be searched and so refer to the previous search's score. // to be searched and so refer to the previous search's score.
if ((value > alpha && value < beta) || current_search_time() > 5000) if ((value > alpha && value < beta) || current_search_time() > 2000)
for (int i = 0; i < Min(UCIMultiPV, (int)Rml.size()); i++) for (int i = 0; i < Min(UCIMultiPV, (int)Rml.size()); i++)
{ {
bool updated = (i <= MultiPVIteration); bool updated = (i <= MultiPVIdx);
if (depth == 1 && !updated) if (depth == 1 && !updated)
continue; continue;
@ -600,14 +604,14 @@ namespace {
cout << "info" cout << "info"
<< depth_to_uci(d) << depth_to_uci(d)
<< (i == MultiPVIteration ? score_to_uci(s, alpha, beta) : score_to_uci(s)) << (i == MultiPVIdx ? score_to_uci(s, alpha, beta) : score_to_uci(s))
<< speed_to_uci(pos.nodes_searched()) << speed_to_uci(pos.nodes_searched())
<< pv_to_uci(&Rml[i].pv[0], i + 1, pos.is_chess960()) << pv_to_uci(&Rml[i].pv[0], i + 1, pos.is_chess960())
<< endl; << endl;
} }
// In case of failing high/low increase aspiration window and research, // In case of failing high/low increase aspiration window and
// otherwise exit the fail high/low loop. // research, otherwise exit the fail high/low loop.
if (value >= beta) if (value >= beta)
{ {
beta = Min(beta + aspirationDelta, VALUE_INFINITE); beta = Min(beta + aspirationDelta, VALUE_INFINITE);
@ -633,14 +637,14 @@ namespace {
bestValues[depth] = value; bestValues[depth] = value;
bestMoveChanges[depth] = Rml.bestMoveChanges; bestMoveChanges[depth] = Rml.bestMoveChanges;
// Do we need to pick now the best and the ponder moves ? // Skills: Do we need to pick now the best and the ponder moves ?
if (SkillLevelEnabled && depth == 1 + SkillLevel) if (SkillLevelEnabled && depth == 1 + SkillLevel)
do_skill_level(&skillBest, &skillPonder); do_skill_level(&skillBest, &skillPonder);
if (LogFile.is_open()) if (LogFile.is_open())
LogFile << pretty_pv(pos, depth, value, current_search_time(), &Rml[0].pv[0]) << endl; LogFile << pretty_pv(pos, depth, value, current_search_time(), &Rml[0].pv[0]) << endl;
// Init easyMove after first iteration or drop if differs from the best move // Init easyMove at first iteration or drop it if differs from the best move
if (depth == 1 && (Rml.size() == 1 || Rml[0].score > Rml[1].score + EasyMoveMargin)) if (depth == 1 && (Rml.size() == 1 || Rml[0].score > Rml[1].score + EasyMoveMargin))
easyMove = bestMove; easyMove = bestMove;
else if (bestMove != easyMove) else if (bestMove != easyMove)
@ -649,9 +653,9 @@ namespace {
// Check for some early stop condition // Check for some early stop condition
if (!StopRequest && Limits.useTimeManagement()) if (!StopRequest && Limits.useTimeManagement())
{ {
// Stop search early if one move seems to be much better than the // Easy move: Stop search early if one move seems to be much better
// others or if there is only a single legal move. Also in the latter // than the others or if there is only a single legal move. Also in
// case we search up to some depth anyway to get a proper score. // the latter case search to some depth anyway to get a proper score.
if ( depth >= 7 if ( depth >= 7
&& easyMove == bestMove && easyMove == bestMove
&& ( Rml.size() == 1 && ( Rml.size() == 1
@ -779,7 +783,7 @@ namespace {
excludedMove = ss->excludedMove; excludedMove = ss->excludedMove;
posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key(); posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key();
tte = TT.probe(posKey); tte = TT.probe(posKey);
ttMove = RootNode ? Rml[MultiPVIteration].pv[0] : tte ? tte->move() : MOVE_NONE; ttMove = RootNode ? Rml[MultiPVIdx].pv[0] : tte ? tte->move() : MOVE_NONE;
// At PV nodes we check for exact scores, while at non-PV nodes we check for // At PV nodes we check for exact scores, while at non-PV nodes we check for
// a fail high/low. Biggest advantage at probing at PV nodes is to have a // a fail high/low. Biggest advantage at probing at PV nodes is to have a
@ -988,10 +992,10 @@ split_point_start: // At split points actual search starts from here
if (move == excludedMove) if (move == excludedMove)
continue; continue;
// At root obey the "searchmoves" option and skip moves not listed in Root Move List. // At root obey the "searchmoves" option and skip moves not listed in Root
// Also in MultiPV mode we skip moves which already have got an exact score // Move List, as a consequence any illegal move is also skipped. In MultiPV
// in previous MultiPV Iteration. Finally any illegal move is skipped here. // mode we also skip PV moves which have been already searched.
if (RootNode && !Rml.find(move, MultiPVIteration)) if (RootNode && !Rml.find(move, MultiPVIdx))
continue; continue;
// At PV and SpNode nodes we want all moves to be legal since the beginning // At PV and SpNode nodes we want all moves to be legal since the beginning
@ -1018,7 +1022,7 @@ split_point_start: // At split points actual search starts from here
if (pos.thread() == 0 && current_search_time() > 2000) if (pos.thread() == 0 && current_search_time() > 2000)
cout << "info" << depth_to_uci(depth) cout << "info" << depth_to_uci(depth)
<< " currmove " << move << " currmove " << move
<< " currmovenumber " << moveCount + MultiPVIteration << endl; << " currmovenumber " << moveCount + MultiPVIdx << endl;
} }
// At Root and at first iteration do a PV search on all the moves to score root moves // At Root and at first iteration do a PV search on all the moves to score root moves