size_t DfpnSolver::MID(const DfpnBounds& maxBounds, DfpnHistory& history)
{
maxBounds.CheckConsistency();
SG_ASSERT(maxBounds.phi > 1);
SG_ASSERT(maxBounds.delta > 1);
++m_numMIDcalls;
size_t prevWork = 0;
SgEmptyBlackWhite colorToMove = GetColorToMove();
DfpnData data;
if (TTRead(data))
{
prevWork = data.m_work;
if (! maxBounds.GreaterThan(data.m_bounds))
// Estimated bounds are larger than we had
// anticipated. The calling state must have computed
// the max bounds with out of date information, so just
// return here without doing anything: the caller will
// now update to this new info and carry on.
return 0;
}
else
{
SgEmptyBlackWhite winner = SG_EMPTY;
if (TerminalState(colorToMove, winner))
{
++m_numTerminal;
DfpnBounds terminal;
if (colorToMove == winner)
DfpnBounds::SetToWinning(terminal);
else
{
SG_ASSERT(SgOppBW(colorToMove) == winner);
DfpnBounds::SetToLosing(terminal);
}
TTWrite(DfpnData(terminal, SG_NULLMOVE, 1));
return 1;
}
}
++m_generateMoves;
DfpnChildren children;
GenerateChildren(children.Children());
// Not thread safe: perhaps move into while loop below later...
std::vector<DfpnData> childrenData(children.Size());
for (size_t i = 0; i < children.Size(); ++i)
LookupData(childrenData[i], children, i);
// Index used for progressive widening
size_t maxChildIndex = ComputeMaxChildIndex(childrenData);
SgHashCode currentHash = Hash();
SgMove bestMove = SG_NULLMOVE;
DfpnBounds currentBounds;
size_t localWork = 1;
do
{
UpdateBounds(currentBounds, childrenData, maxChildIndex);
if (! maxBounds.GreaterThan(currentBounds))
break;
// Select most proving child
std::size_t bestIndex = 999999;
DfpnBoundType delta2 = DfpnBounds::INFTY;
SelectChild(bestIndex, delta2, childrenData, maxChildIndex);
bestMove = children.MoveAt(bestIndex);
// Compute maximum bound for child
const DfpnBounds childBounds(childrenData[bestIndex].m_bounds);
DfpnBounds childMaxBounds;
childMaxBounds.phi = maxBounds.delta
- (currentBounds.delta - childBounds.phi);
childMaxBounds.delta = delta2 == DfpnBounds::INFTY ? maxBounds.phi :
std::min(maxBounds.phi,
std::max(delta2 + 1, DfpnBoundType(delta2 * (1.0 + m_epsilon))));
SG_ASSERT(childMaxBounds.GreaterThan(childBounds));
if (delta2 != DfpnBounds::INFTY)
m_deltaIncrease.Add(float(childMaxBounds.delta-childBounds.delta));
// Recurse on best child
PlayMove(bestMove);
history.Push(bestMove, currentHash);
localWork += MID(childMaxBounds, history);
history.Pop();
UndoMove();
// Update bounds for best child
LookupData(childrenData[bestIndex], children, bestIndex);
// Compute some stats when find winning move
if (childrenData[bestIndex].m_bounds.IsLosing())
{
m_moveOrderingIndex.Add(float(bestIndex));
m_moveOrderingPercent.Add(float(bestIndex)
/ (float)childrenData.size());
m_totalWastedWork += prevWork + localWork
- childrenData[bestIndex].m_work;
}
else if (childrenData[bestIndex].m_bounds.IsWinning())
maxChildIndex = ComputeMaxChildIndex(childrenData);
} while (! CheckAbort());
// Find the most delaying move for losing states, and the smallest
// winning move for winning states.
if (currentBounds.IsSolved())
{
if (currentBounds.IsLosing())
{
std::size_t maxWork = 0;
for (std::size_t i = 0; i < children.Size(); ++i)
{
if (childrenData[i].m_work > maxWork)
{
maxWork = childrenData[i].m_work;
bestMove = children.MoveAt(i);
}
}
}
else
{
std::size_t minWork = DfpnBounds::INFTY;
for (std::size_t i = 0; i < children.Size(); ++i)
{
if (childrenData[i].m_bounds.IsLosing()
&& childrenData[i].m_work < minWork)
{
minWork = childrenData[i].m_work;
bestMove = children.MoveAt(i);
}
}
}
}
// Store search results
TTWrite(DfpnData(currentBounds, bestMove, localWork + prevWork));
return localWork;
}
size_t DfpnSolver::MID(const DfpnBounds& maxBounds, DfpnHistory& history)
{
maxBounds.CheckConsistency();
HexAssert(maxBounds.phi > 1);
HexAssert(maxBounds.delta > 1);
int depth = history.Depth();
size_t prevWork = 0;
bitset_t maxProofSet;
float evaluationScore;
HexColor colorToMove = m_state->ToPlay();
DfpnChildren children;
{
DfpnData data;
if (TTRead(*m_state, data))
{
children = data.m_children;
maxProofSet = data.m_maxProofSet;
prevWork = data.m_work;
evaluationScore = data.m_evaluationScore;
if (!maxBounds.GreaterThan(data.m_bounds))
// Estimated bounds are larger than we had
// anticipated. The calling state must have computed
// the max bounds with out of date information, so just
// return here without doing anything: the caller will
// now update to this new info and carry on.
return 0;
}
else
{
m_workBoard->GetPosition().SetPosition(m_state->Position());
m_workBoard->ComputeAll(colorToMove);
++m_numVCbuilds;
// Compute the maximum possible proof set if colorToMove wins.
// This data is used to prune siblings of this state.
maxProofSet = ProofUtil::MaximumProofSet(*m_workBoard, colorToMove);
if (EndgameUtils::IsDeterminedState(*m_workBoard, colorToMove))
{
++m_numTerminal;
DfpnBounds terminal;
if (EndgameUtils::IsWonGame(*m_workBoard, colorToMove))
DfpnBounds::SetToWinning(terminal);
else
DfpnBounds::SetToLosing(terminal);
if (m_useGuiFx && depth == 1)
{
m_guiFx.UpdateCurrentBounds(terminal);
m_guiFx.Write();
}
TTWrite(*m_state, DfpnData(terminal, DfpnChildren(),
INVALID_POINT, 1, maxProofSet, 0.0));
return 1;
}
bitset_t childrenBitset
= EndgameUtils::MovesToConsider(*m_workBoard, colorToMove);
m_considerSetSize.Add(childrenBitset.count());
Resistance resist;
resist.Evaluate(*m_workBoard);
evaluationScore = (colorToMove == BLACK)
? resist.Score() : -resist.Score();
m_allEvaluation.Add(evaluationScore);
std::vector<std::pair<HexEval, HexPoint> > mvsc;
for (BitsetIterator it(childrenBitset); it; ++it)
{
HexEval score = resist.Score(*it);
mvsc.push_back(std::make_pair(-score, *it));
}
stable_sort(mvsc.begin(), mvsc.end());
std::vector<HexPoint> sortedChildren;
for (size_t i = 0; i < mvsc.size(); ++i)
sortedChildren.push_back(mvsc[i].second);
children.SetChildren(sortedChildren);
}
}
++m_numMIDcalls;
size_t localWork = 1;
// Not thread safe: perhaps move into while loop below later...
std::vector<DfpnData> childrenData(children.Size());
for (size_t i = 0; i < children.Size(); ++i)
LookupData(childrenData[i], children, i, *m_state);
// Index used for progressive widening
size_t maxChildIndex = ComputeMaxChildIndex(childrenData);
if (m_useGuiFx && depth == 0)
m_guiFx.SetChildren(children, childrenData);
hash_t currentHash = m_state->Hash();
HexPoint bestMove = INVALID_POINT;
DfpnBounds currentBounds;
do
{
UpdateBounds(currentBounds, childrenData, maxChildIndex);
if (m_useGuiFx && depth == 1)
{
m_guiFx.UpdateCurrentBounds(currentBounds);
m_guiFx.Write();
}
if (!maxBounds.GreaterThan(currentBounds))
break;
// Select most proving child
int bestIndex = -1;
DfpnBoundType delta2 = DfpnBounds::INFTY;
SelectChild(bestIndex, delta2, childrenData, maxChildIndex);
bestMove = children.FirstMove(bestIndex);
// Compute maximum bound for child
const DfpnBounds childBounds(childrenData[bestIndex].m_bounds);
DfpnBounds childMaxBounds;
childMaxBounds.phi = maxBounds.delta
- (currentBounds.delta - childBounds.phi);
childMaxBounds.delta = std::min(maxBounds.phi, delta2 + 1);
HexAssert(childMaxBounds.GreaterThan(childBounds));
if (delta2 != DfpnBounds::INFTY)
m_deltaIncrease.Add(childMaxBounds.delta - childBounds.delta);
// Recurse on best child
if (m_useGuiFx && depth == 0)
m_guiFx.PlayMove(colorToMove, bestIndex);
children.PlayMove(bestIndex, *m_state);
history.Push(bestMove, currentHash);
localWork += MID(childMaxBounds, history);
history.Pop();
children.UndoMove(bestIndex, *m_state);
if (m_useGuiFx && depth == 0)
m_guiFx.UndoMove();
// Update bounds for best child
LookupData(childrenData[bestIndex], children, bestIndex, *m_state);
// Compute some stats when find winning move
if (childrenData[bestIndex].m_bounds.IsLosing())
{
m_moveOrderingIndex.Add(bestIndex);
m_moveOrderingPercent.Add(bestIndex / (double)childrenData.size());
m_totalWastedWork += prevWork + localWork
- childrenData[bestIndex].m_work;
}
else if (childrenData[bestIndex].m_bounds.IsWinning())
maxChildIndex = ComputeMaxChildIndex(childrenData);
// Shrink children list using knowledge of bestMove child's proof set.
// That is, if this child is losing, conclude what other children
// must also be losing (i.e. cannot interfere with the proof set
// that disproves this child).
// And of course if this child is winning, no need to explore
// these other siblings either.
{
/* @todo Perhaps track allChildren instead of recomputing? */
bitset_t allChildren;
for (std::vector<HexPoint>::iterator it
= children.m_children.begin();
it != children.m_children.end(); ++it)
{
allChildren.set(*it);
}
bitset_t canPrune = allChildren
- childrenData[bestIndex].m_maxProofSet;
canPrune.reset(bestMove);
int pruneCount = canPrune.count();
if (pruneCount)
{
m_prunedSiblingStats.Add(pruneCount);
/*
LogInfo() << "Pruning " << pruneCount
<< " moves via " << bestMove
<< ".\nChildren:\n" << m_brd->Write(allChildren)
<< "\nRemoving...\n" << m_brd->Write(canPrune)
<< "\n";
*/
DeleteChildren(children, childrenData, canPrune);
maxChildIndex = ComputeMaxChildIndex(childrenData);
if (m_useGuiFx && depth == 0)
m_guiFx.SetChildren(children, childrenData);
}
}
} while (!CheckAbort());
if (m_useGuiFx && depth == 0)
m_guiFx.WriteForced();
// Find the most delaying move for losing states, and the smallest
// winning move for winning states.
if (currentBounds.IsSolved())
{
m_allSolvedEvaluation.Add(evaluationScore);
if (currentBounds.IsLosing())
{
m_losingEvaluation.Add(evaluationScore);
std::size_t maxWork = 0;
for (std::size_t i = 0; i < children.Size(); ++i)
{
if (childrenData[i].m_work > maxWork)
{
maxWork = childrenData[i].m_work;
bestMove = children.FirstMove(i);
}
}
}
else
{
m_winningEvaluation.Add(evaluationScore);
std::size_t minWork = DfpnBounds::INFTY;
for (std::size_t i = 0; i < children.Size(); ++i)
{
if (childrenData[i].m_bounds.IsLosing()
&& childrenData[i].m_work < minWork)
{
minWork = childrenData[i].m_work;
bestMove = children.FirstMove(i);
}
}
}
}
// Store search results and notify listeners
DfpnData data(currentBounds, children, bestMove, localWork + prevWork,
maxProofSet, evaluationScore);
TTWrite(*m_state, data);
if (data.m_bounds.IsSolved())
NotifyListeners(history, data);
return localWork;
}
