Add Human-SL KGS-rank ladder (gtp_human<rank>.cfg, 9d→20k) + tunehuman subcommand

cpp/CMakeLists.txt

@@ -321,6 +321,7 @@ add_executable(katago
   program/playutils.cpp
   program/playsettings.cpp
   program/play.cpp
+  program/humansltuner.cpp
   program/selfplaymanager.cpp
   ${GIT_HEADER_FILE_ALWAYS_UPDATED}
   tests/testboardarea.cpp
@@ -333,6 +334,7 @@ add_executable(katago
   tests/testrules.cpp
   tests/testscore.cpp
   tests/testsgf.cpp
+  tests/testhumansltuner.cpp
   tests/testsymmetries.cpp
   tests/testnninputs.cpp
   tests/testownership.cpp
@@ -365,6 +367,7 @@ add_executable(katago
   command/selfplay.cpp
   command/startposes.cpp
   command/tune.cpp
+  command/tunehuman.cpp
   command/writetrainingdata.cpp
   main.cpp
   )

cpp/command/runtests.cpp

@@ -55,6 +55,8 @@ int MainCmds::runtests(const vector<string>& args) {
   Tests::runSymmetryDifferenceTests();
   Tests::runBoardReplayTest();
 
+  Tests::runHumanSLTunerTests();
+
   ScoreValue::freeTables();
 
   Tests::runInlineConfigTests();

cpp/configs/gtp_human10k.cfg

@@ -0,0 +1,74 @@
+# CALIBRATED (Japanese, komi-0.5 handicap, ANE): preaz_10k (Black) vs gtp_human9k.cfg (White) = 50.0% [42.0,58.0] over 148 games. λ=0.59036. LOCK.
+# gtp_human10k.cfg — 10k rung of the Human-SL KGS-rank ladder. See docs/HumanSL_Rank_Ladder.md.
+# Tuned so preaz_10k (Black, komi 0.5) is an even game (50%) vs the prior rung gtp_human9k.cfg (White)
+# = exactly 1 KGS rank (1 stone) weaker. Strength dial = humanSLChosenMovePiklLambda (below); 400 visits.
+#
+# Run: ./katago gtp -config gtp_human10k.cfg -model <katago_model>.bin.gz -human-model b18c384nbt-humanv0.bin.gz
+logDir = gtp_logs
+logAllGTPCommunication = true
+logSearchInfo = true
+logSearchInfoForChosenMove = false
+logToStderr = false
+
+rules = japanese
+
+allowResignation = true
+resignThreshold = -0.98
+resignConsecTurns = 10
+resignMinScoreDifference = 20
+resignMinMovesPerBoardArea = 0.40
+
+# This ladder runs at a fixed 400 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 400
+
+numSearchThreads = 8
+lagBuffer = 1.0
+
+delayMoveScale = 2.0
+delayMoveMax = 10.0
+
+# Imitate a human amateur at this rung KGS rank (Human-SL profile set in humanSLProfile below).
+humanSLProfile = preaz_10k
+humanSLChosenMoveProp = 1.0
+humanSLChosenMoveIgnorePass = true
+
+# Strength dial: higher humanSLChosenMovePiklLambda is more human and weaker; lower trusts search and is stronger.
+# The calibrated value for this rung is set below; see the CALIBRATED line at the top of this file.
+# Higher -> more human / weaker; lower -> trusts KataGo search more / stronger.
+humanSLChosenMovePiklLambda = 0.59036
+
+# Spend 80% of visits to explore humanSL moves so they get evaluations for humanSLChosenMovePiklLambda
+humanSLRootExploreProbWeightless = 0.8
+humanSLRootExploreProbWeightful = 0.0
+humanSLPlaExploreProbWeightless = 0.0
+humanSLPlaExploreProbWeightful = 0.0
+humanSLOppExploreProbWeightless = 0.0
+humanSLOppExploreProbWeightful = 0.0
+
+humanSLCpuctExploration = 0.50
+humanSLCpuctPermanent = 2.0
+
+chosenMoveTemperatureEarly = 0.70
+chosenMoveTemperature = 0.25
+chosenMoveTemperatureHalflife = 30
+chosenMoveTemperatureOnlyBelowProb = 1.0
+chosenMoveSubtract = 0
+chosenMovePrune = 0
+
+nnCacheSizePowerOfTwo = 20
+nnMutexPoolSizePowerOfTwo = 14
+
+ignorePreRootHistory = false
+analysisIgnorePreRootHistory = false
+
+rootNumSymmetriesToSample = 2
+useLcbForSelection = false
+
+winLossUtilityFactor = 1.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human11k.cfg

@@ -0,0 +1,74 @@
+# CALIBRATED (Japanese, komi-0.5 handicap, ANE): preaz_11k (Black) vs gtp_human10k.cfg (White) = 48.1% [40.5,55.8] over 160 games. λ=0.56458. LOCK.
+# gtp_human11k.cfg — 11k rung of the Human-SL KGS-rank ladder. See docs/HumanSL_Rank_Ladder.md.
+# Tuned so preaz_11k (Black, komi 0.5) is an even game (50%) vs the prior rung gtp_human10k.cfg (White)
+# = exactly 1 KGS rank (1 stone) weaker. Strength dial = humanSLChosenMovePiklLambda (below); 400 visits.
+#
+# Run: ./katago gtp -config gtp_human11k.cfg -model <katago_model>.bin.gz -human-model b18c384nbt-humanv0.bin.gz
+logDir = gtp_logs
+logAllGTPCommunication = true
+logSearchInfo = true
+logSearchInfoForChosenMove = false
+logToStderr = false
+
+rules = japanese
+
+allowResignation = true
+resignThreshold = -0.98
+resignConsecTurns = 10
+resignMinScoreDifference = 20
+resignMinMovesPerBoardArea = 0.40
+
+# This ladder runs at a fixed 400 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 400
+
+numSearchThreads = 8
+lagBuffer = 1.0
+
+delayMoveScale = 2.0
+delayMoveMax = 10.0
+
+# Imitate a human amateur at this rung KGS rank (Human-SL profile set in humanSLProfile below).
+humanSLProfile = preaz_11k
+humanSLChosenMoveProp = 1.0
+humanSLChosenMoveIgnorePass = true
+
+# Strength dial: higher humanSLChosenMovePiklLambda is more human and weaker; lower trusts search and is stronger.
+# The calibrated value for this rung is set below; see the CALIBRATED line at the top of this file.
+# Higher -> more human / weaker; lower -> trusts KataGo search more / stronger.
+humanSLChosenMovePiklLambda = 0.56458
+
+# Spend 80% of visits to explore humanSL moves so they get evaluations for humanSLChosenMovePiklLambda
+humanSLRootExploreProbWeightless = 0.8
+humanSLRootExploreProbWeightful = 0.0
+humanSLPlaExploreProbWeightless = 0.0
+humanSLPlaExploreProbWeightful = 0.0
+humanSLOppExploreProbWeightless = 0.0
+humanSLOppExploreProbWeightful = 0.0
+
+humanSLCpuctExploration = 0.50
+humanSLCpuctPermanent = 2.0
+
+chosenMoveTemperatureEarly = 0.70
+chosenMoveTemperature = 0.25
+chosenMoveTemperatureHalflife = 30
+chosenMoveTemperatureOnlyBelowProb = 1.0
+chosenMoveSubtract = 0
+chosenMovePrune = 0
+
+nnCacheSizePowerOfTwo = 20
+nnMutexPoolSizePowerOfTwo = 14
+
+ignorePreRootHistory = false
+analysisIgnorePreRootHistory = false
+
+rootNumSymmetriesToSample = 2
+useLcbForSelection = false
+
+winLossUtilityFactor = 1.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human12k.cfg

@@ -0,0 +1,74 @@
+# CALIBRATED (Japanese, komi-0.5 handicap, ANE): preaz_12k (Black) vs gtp_human11k.cfg (White) = 50.8% [42.2,59.3] over 128 games. λ=0.54297. LOCK.
+# gtp_human12k.cfg — 12k rung of the Human-SL KGS-rank ladder. See docs/HumanSL_Rank_Ladder.md.
+# Tuned so preaz_12k (Black, komi 0.5) is an even game (50%) vs the prior rung gtp_human11k.cfg (White)
+# = exactly 1 KGS rank (1 stone) weaker. Strength dial = humanSLChosenMovePiklLambda (below); 400 visits.
+#
+# Run: ./katago gtp -config gtp_human12k.cfg -model <katago_model>.bin.gz -human-model b18c384nbt-humanv0.bin.gz
+logDir = gtp_logs
+logAllGTPCommunication = true
+logSearchInfo = true
+logSearchInfoForChosenMove = false
+logToStderr = false
+
+rules = japanese
+
+allowResignation = true
+resignThreshold = -0.98
+resignConsecTurns = 10
+resignMinScoreDifference = 20
+resignMinMovesPerBoardArea = 0.40
+
+# This ladder runs at a fixed 400 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 400
+
+numSearchThreads = 8
+lagBuffer = 1.0
+
+delayMoveScale = 2.0
+delayMoveMax = 10.0
+
+# Imitate a human amateur at this rung KGS rank (Human-SL profile set in humanSLProfile below).
+humanSLProfile = preaz_12k
+humanSLChosenMoveProp = 1.0
+humanSLChosenMoveIgnorePass = true
+
+# Strength dial: higher humanSLChosenMovePiklLambda is more human and weaker; lower trusts search and is stronger.
+# The calibrated value for this rung is set below; see the CALIBRATED line at the top of this file.
+# Higher -> more human / weaker; lower -> trusts KataGo search more / stronger.
+humanSLChosenMovePiklLambda = 0.54297
+
+# Spend 80% of visits to explore humanSL moves so they get evaluations for humanSLChosenMovePiklLambda
+humanSLRootExploreProbWeightless = 0.8
+humanSLRootExploreProbWeightful = 0.0
+humanSLPlaExploreProbWeightless = 0.0
+humanSLPlaExploreProbWeightful = 0.0
+humanSLOppExploreProbWeightless = 0.0
+humanSLOppExploreProbWeightful = 0.0
+
+humanSLCpuctExploration = 0.50
+humanSLCpuctPermanent = 2.0
+
+chosenMoveTemperatureEarly = 0.70
+chosenMoveTemperature = 0.25
+chosenMoveTemperatureHalflife = 30
+chosenMoveTemperatureOnlyBelowProb = 1.0
+chosenMoveSubtract = 0
+chosenMovePrune = 0
+
+nnCacheSizePowerOfTwo = 20
+nnMutexPoolSizePowerOfTwo = 14
+
+ignorePreRootHistory = false
+analysisIgnorePreRootHistory = false
+
+rootNumSymmetriesToSample = 2
+useLcbForSelection = false
+
+winLossUtilityFactor = 1.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human13k.cfg

@@ -0,0 +1,74 @@
+# CALIBRATED (Japanese, komi-0.5 handicap, ANE): preaz_13k (Black) vs gtp_human12k.cfg (White) = 50.8% [42.1,59.4] over 124 games. λ=0.58977. LOCK.
+# gtp_human13k.cfg — 13k rung of the Human-SL KGS-rank ladder. See docs/HumanSL_Rank_Ladder.md.
+# Tuned so preaz_13k (Black, komi 0.5) is an even game (50%) vs the prior rung gtp_human12k.cfg (White)
+# = exactly 1 KGS rank (1 stone) weaker. Strength dial = humanSLChosenMovePiklLambda (below); 400 visits.
+#
+# Run: ./katago gtp -config gtp_human13k.cfg -model <katago_model>.bin.gz -human-model b18c384nbt-humanv0.bin.gz
+logDir = gtp_logs
+logAllGTPCommunication = true
+logSearchInfo = true
+logSearchInfoForChosenMove = false
+logToStderr = false
+
+rules = japanese
+
+allowResignation = true
+resignThreshold = -0.98
+resignConsecTurns = 10
+resignMinScoreDifference = 20
+resignMinMovesPerBoardArea = 0.40
+
+# This ladder runs at a fixed 400 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 400
+
+numSearchThreads = 8
+lagBuffer = 1.0
+
+delayMoveScale = 2.0
+delayMoveMax = 10.0
+
+# Imitate a human amateur at this rung KGS rank (Human-SL profile set in humanSLProfile below).
+humanSLProfile = preaz_13k
+humanSLChosenMoveProp = 1.0
+humanSLChosenMoveIgnorePass = true
+
+# Strength dial: higher humanSLChosenMovePiklLambda is more human and weaker; lower trusts search and is stronger.
+# The calibrated value for this rung is set below; see the CALIBRATED line at the top of this file.
+# Higher -> more human / weaker; lower -> trusts KataGo search more / stronger.
+humanSLChosenMovePiklLambda = 0.58977
+
+# Spend 80% of visits to explore humanSL moves so they get evaluations for humanSLChosenMovePiklLambda
+humanSLRootExploreProbWeightless = 0.8
+humanSLRootExploreProbWeightful = 0.0
+humanSLPlaExploreProbWeightless = 0.0
+humanSLPlaExploreProbWeightful = 0.0
+humanSLOppExploreProbWeightless = 0.0
+humanSLOppExploreProbWeightful = 0.0
+
+humanSLCpuctExploration = 0.50
+humanSLCpuctPermanent = 2.0
+
+chosenMoveTemperatureEarly = 0.70
+chosenMoveTemperature = 0.25
+chosenMoveTemperatureHalflife = 30
+chosenMoveTemperatureOnlyBelowProb = 1.0
+chosenMoveSubtract = 0
+chosenMovePrune = 0
+
+nnCacheSizePowerOfTwo = 20
+nnMutexPoolSizePowerOfTwo = 14
+
+ignorePreRootHistory = false
+analysisIgnorePreRootHistory = false
+
+rootNumSymmetriesToSample = 2
+useLcbForSelection = false
+
+winLossUtilityFactor = 1.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false