Simplify Skill implementation
Currently we handle the UCI_Elo with a double randomization. This seems not necessary and a bit involuted. This patch removes the first randomization and unifies the 2 cases. closes https://github.com/official-stockfish/Stockfish/pull/3769 No functional change.pull/3764/head^2
parent
0e89d6e754
commit
ef4822aa8d
1
AUTHORS
1
AUTHORS
|
@ -194,6 +194,7 @@ tttak
|
|||
Unai Corzo (unaiic)
|
||||
Uri Blass (uriblass)
|
||||
Vince Negri (cuddlestmonkey)
|
||||
xefoci7612
|
||||
zz4032
|
||||
|
||||
|
||||
|
|
|
@ -112,14 +112,22 @@ namespace {
|
|||
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;
|
||||
};
|
||||
|
||||
|
@ -243,10 +251,11 @@ void MainThread::search() {
|
|||
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();
|
||||
|
||||
|
@ -311,19 +320,7 @@ void Thread::search() {
|
|||
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.
|
||||
|
@ -1780,8 +1777,8 @@ moves_loop: // When in check, search starts here
|
|||
// 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
|
||||
|
@ -1789,8 +1786,8 @@ moves_loop: // When in check, search starts here
|
|||
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)
|
||||
{
|
||||
|
|
Loading…
Reference in New Issue