return thisThread->state;
}
- // Skill structure is used to implement strength limit
+ // Skill structure is used to implement strength limit. If we have an uci_elo then
+ // we convert it to a suitable fractional skill level using anchoring to CCRL Elo
+ // (goldfish 1.13 = 2000) and a fit through Ordo derived Elo for match (TC 60+0.6)
+ // results spanning a wide range of k values.
struct Skill {
- explicit Skill(int l) : level(l) {}
- bool enabled() const { return level < 20; }
- bool time_to_pick(Depth depth) const { return depth == 1 + level; }
+ Skill(int skill_level, int uci_elo) {
+ if (uci_elo)
+ level = std::clamp(std::pow((uci_elo - 1346.6) / 143.4, 1 / 0.806), 0.0, 20.0);
+ else
+ level = double(skill_level);
+ }
+ bool enabled() const { return level < 20.0; }
+ bool time_to_pick(Depth depth) const { return depth == 1 + int(level); }
Move pick_best(size_t multiPV);
- int level;
+ double level;
Move best = MOVE_NONE;
};
Time.availableNodes += Limits.inc[us] - Threads.nodes_searched();
Thread* bestThread = this;
+ Skill skill = Skill(Options["Skill Level"], Options["UCI_LimitStrength"] ? int(Options["UCI_Elo"]) : 0);
if ( int(Options["MultiPV"]) == 1
&& !Limits.depth
- && !(Skill(Options["Skill Level"]).enabled() || int(Options["UCI_LimitStrength"]))
+ && !skill.enabled()
&& rootMoves[0].pv[0] != MOVE_NONE)
bestThread = Threads.get_best_thread();
std::fill(&lowPlyHistory[MAX_LPH - 2][0], &lowPlyHistory.back().back() + 1, 0);
size_t multiPV = size_t(Options["MultiPV"]);
-
- // Pick integer skill levels, but non-deterministically round up or down
- // such that the average integer skill corresponds to the input floating point one.
- // UCI_Elo is converted to a suitable fractional skill level, using anchoring
- // to CCRL Elo (goldfish 1.13 = 2000) and a fit through Ordo derived Elo
- // for match (TC 60+0.6) results spanning a wide range of k values.
- PRNG rng(now());
- double floatLevel = Options["UCI_LimitStrength"] ?
- std::clamp(std::pow((Options["UCI_Elo"] - 1346.6) / 143.4, 1 / 0.806), 0.0, 20.0) :
- double(Options["Skill Level"]);
- int intLevel = int(floatLevel) +
- ((floatLevel - int(floatLevel)) * 1024 > rng.rand<unsigned>() % 1024 ? 1 : 0);
- Skill skill(intLevel);
+ Skill skill(Options["Skill Level"], Options["UCI_LimitStrength"] ? int(Options["UCI_Elo"]) : 0);
// When playing with strength handicap enable MultiPV search that we will
// use behind the scenes to retrieve a set of possible moves.
if ( !PvNode
&& depth < 9
&& eval - futility_margin(depth, improving) >= beta
- && eval < VALUE_KNOWN_WIN) // Do not return unproven wins
+ && eval < 15000) // 50% larger than VALUE_KNOWN_WIN, but smaller than TB wins.
return eval;
// Step 8. Null move search with verification search (~40 Elo)
// In general we want to cap the LMR depth search at newDepth. But if reductions
// are really negative and movecount is low, we allow this move to be searched
// deeper than the first move (this may lead to hidden double extensions).
- int deeper = r >= -1 ? 0
- : moveCount <= 5 ? 2
- : PvNode && depth > 6 ? 1
- : 0;
+ int deeper = r >= -1 ? 0
+ : moveCount <= 5 ? 2
+ : PvNode && depth > 6 ? 1
+ : cutNode && moveCount <= 7 ? 1
+ : 0;
Depth d = std::clamp(newDepth - r, 1, newDepth + deeper);
// RootMoves are already sorted by score in descending order
Value topScore = rootMoves[0].score;
int delta = std::min(topScore - rootMoves[multiPV - 1].score, PawnValueMg);
- int weakness = 120 - 2 * level;
int maxScore = -VALUE_INFINITE;
+ double weakness = 120 - 2 * level;
// Choose best move. For each move score we add two terms, both dependent on
// weakness. One is deterministic and bigger for weaker levels, and one is
for (size_t i = 0; i < multiPV; ++i)
{
// This is our magic formula
- int push = ( weakness * int(topScore - rootMoves[i].score)
- + delta * (rng.rand<unsigned>() % weakness)) / 128;
+ int push = int(( weakness * int(topScore - rootMoves[i].score)
+ + delta * (rng.rand<unsigned>() % int(weakness))) / 128);
if (rootMoves[i].score + push >= maxScore)
{