Add Human-SL even-game ELO rank ladder configs (gtp_human<rank>.cfg, 8d→25k)

cpp/configs/gtp_human10k.cfg

@@ -0,0 +1,71 @@
+# CALIBRATED (Japanese, even-game ELO ladder (komi 6.5, alternating colors; 40v; winLossUtilityFactor=0; b28c512 main net)): preaz_10k vs gtp_human9k.cfg, even game = 35.9% -> gap +101 ELO [74,128] over 688 games (target 100, 95% CI ⊂ [70,130]). λ=0.37250.
+# gtp_human10k.cfg — 10k rung of the Human-SL even-game ELO ladder. See docs/HumanSL_Rank_Ladder.md.
+# Run: ./katago gtp -config gtp_human10k.cfg -model kata1-b28c512nbt-s8326494464-d4628051565.bin.gz -human-model b18c384nbt-humanv0.bin.gz   (calibrated for THIS main net; a different one invalidates the calibration)
+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 40 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 40
+
+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 = 0.37250
+
+# 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 = 0.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human11k.cfg

@@ -0,0 +1,71 @@
+# CALIBRATED (Japanese, even-game ELO ladder (komi 6.5, alternating colors; 40v; winLossUtilityFactor=0; b28c512 main net)): preaz_11k vs gtp_human10k.cfg, even game = 35.7% -> gap +103 ELO [76,129] over 704 games (target 100, 95% CI ⊂ [70,130]). λ=0.40810.
+# gtp_human11k.cfg — 11k rung of the Human-SL even-game ELO ladder. See docs/HumanSL_Rank_Ladder.md.
+# Run: ./katago gtp -config gtp_human11k.cfg -model kata1-b28c512nbt-s8326494464-d4628051565.bin.gz -human-model b18c384nbt-humanv0.bin.gz   (calibrated for THIS main net; a different one invalidates the calibration)
+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 40 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 40
+
+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 = 0.40810
+
+# 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 = 0.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human12k.cfg

@@ -0,0 +1,71 @@
+# CALIBRATED (Japanese, even-game ELO ladder (komi 6.5, alternating colors; 40v; winLossUtilityFactor=0; b28c512 main net)): preaz_12k vs gtp_human11k.cfg, even game = 36.0% -> gap +100 ELO [71,129] over 608 games (target 100, 95% CI ⊂ [70,130]). λ=0.46300.
+# gtp_human12k.cfg — 12k rung of the Human-SL even-game ELO ladder. See docs/HumanSL_Rank_Ladder.md.
+# Run: ./katago gtp -config gtp_human12k.cfg -model kata1-b28c512nbt-s8326494464-d4628051565.bin.gz -human-model b18c384nbt-humanv0.bin.gz   (calibrated for THIS main net; a different one invalidates the calibration)
+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 40 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 40
+
+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 = 0.46300
+
+# 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 = 0.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human13k.cfg

@@ -0,0 +1,71 @@
+# CALIBRATED (Japanese, even-game ELO ladder (komi 6.5, alternating colors; 40v; winLossUtilityFactor=0; b28c512 main net)): preaz_13k vs gtp_human12k.cfg, even game = 35.5% -> gap +104 ELO [79,129] over 800 games (target 100, 95% CI ⊂ [70,130]). λ=0.83000.
+# gtp_human13k.cfg — 13k rung of the Human-SL even-game ELO ladder. See docs/HumanSL_Rank_Ladder.md.
+# Run: ./katago gtp -config gtp_human13k.cfg -model kata1-b28c512nbt-s8326494464-d4628051565.bin.gz -human-model b18c384nbt-humanv0.bin.gz   (calibrated for THIS main net; a different one invalidates the calibration)
+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 40 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 40
+
+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 = 0.83000
+
+# 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 = 0.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false

cpp/configs/gtp_human14k.cfg

@@ -0,0 +1,71 @@
+# CALIBRATED (Japanese, even-game ELO ladder (komi 6.5, alternating colors; 40v; winLossUtilityFactor=0; b28c512 main net)): preaz_14k vs gtp_human13k.cfg, even game = 36.0% -> gap +100 ELO [71,129] over 592 games (target 100, 95% CI ⊂ [70,130]). λ=3.40040.
+# gtp_human14k.cfg — 14k rung of the Human-SL even-game ELO ladder. See docs/HumanSL_Rank_Ladder.md.
+# Run: ./katago gtp -config gtp_human14k.cfg -model kata1-b28c512nbt-s8326494464-d4628051565.bin.gz -human-model b18c384nbt-humanv0.bin.gz   (calibrated for THIS main net; a different one invalidates the calibration)
+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 40 visits; strength is set by humanSLChosenMovePiklLambda (below), not visits.
+# To adjust strength, change humanSLChosenMovePiklLambda (below), not maxVisits.
+maxVisits = 40
+
+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_14k
+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 = 3.40040
+
+# 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 = 0.0
+staticScoreUtilityFactor = 0.5
+dynamicScoreUtilityFactor = 0.5
+
+useUncertainty = false
+subtreeValueBiasFactor = 0.0
+useNoisePruning = false