From 1484344706aa598f5535c730c2c9ecdfc8907dcb Mon Sep 17 00:00:00 2001
From: farid <farid.mohammadi@iws.uni-stuttgart.de>
Date: Sat, 29 Feb 2020 19:00:24 +0100
Subject: [PATCH] [surrogate] added debug log-likelihood plots.
---
.../surrogate_models/surrogate_models.py | 36 +++++++++++++++----
.../AnalyticalFunction_Test.py | 16 ++++-----
2 files changed, 37 insertions(+), 15 deletions(-)
diff --git a/BayesValidRox/surrogate_models/surrogate_models.py b/BayesValidRox/surrogate_models/surrogate_models.py
index 2d88abcaf..12a6da1da 100644
--- a/BayesValidRox/surrogate_models/surrogate_models.py
+++ b/BayesValidRox/surrogate_models/surrogate_models.py
@@ -250,14 +250,19 @@ class aPCE:
if self.RegMethod != 'AaPCE':
if self.RegMethod == 'BRR':
clf_poly = linear_model.BayesianRidge(n_iter=1000, tol=1e-7,
- fit_intercept = False, compute_score=True)
-
+ fit_intercept = False,
+ compute_score=True,
+ alpha_1=1e-20, alpha_2=1e-20,
+ lambda_1=1e-7, lambda_2=1e-7)
+
+
elif self.RegMethod == 'ARD':
- clf_poly = ARDRegression(fit_intercept = False,compute_score=True)#,
-# n_iter=1000, tol= 1e-7,
-# alpha_1=1e-05, alpha_2=1e-05,
-# lambda_1=1e-9, lambda_2=1e-9)
-
+ clf_poly = ARDRegression(fit_intercept = False,compute_score=True,
+ n_iter=500, tol= 0.001,
+ alpha_1=1e-06, alpha_2=1e-06,
+ lambda_1=1e-1, lambda_2=1e-1)
+
+
elif self.RegMethod == 'LARS':
clf_poly = linear_model.Lars(fit_intercept=False)
@@ -265,6 +270,7 @@ class aPCE:
clf_poly.fit(PSI, ModelOutput)
#print("R^2:\n", clf_poly.score(PSI, ModelOutput))
+
# Select the nonzero entries of coefficients
# The first column must be kept (For mean calculations)
nnz_idx = np.nonzero(clf_poly.coef_)[0]
@@ -404,6 +410,22 @@ class aPCE:
print("Sparsity index:", round(len(PolynomialDegrees)/P, 3))
print("Best Indices:\n", PolynomialDegrees)
+ if self.RegMethod in ['BRR', 'ARD']:
+ plt.figure(figsize=(12, 10))
+ plt.title("Marginal log-likelihood")
+ plt.plot(clf_poly.scores_, color='navy', linewidth=2)
+ plt.ylabel("Score")
+ plt.xlabel("Iterations")
+ try:
+ text = "$\\alpha={:.1f}$\n$\\lambda={:.3f}$\n$L={:.1f}$".format(
+ clf_poly.alpha_, clf_poly.lambda_, clf_poly.scores_[-1])
+ except:
+ text = "$\\alpha={:.1f}$\n$\\$L={:.1f}$".format(
+ clf_poly.alpha_, clf_poly.scores_[-1])
+
+ plt.text(0.05, -1.0, text, fontsize=18)
+ plt.show()
+
print ('='*80)
diff --git a/BayesValidRox/tests/AnalyticalFunction/AnalyticalFunction_Test.py b/BayesValidRox/tests/AnalyticalFunction/AnalyticalFunction_Test.py
index 41e101459..2a74d4539 100644
--- a/BayesValidRox/tests/AnalyticalFunction/AnalyticalFunction_Test.py
+++ b/BayesValidRox/tests/AnalyticalFunction/AnalyticalFunction_Test.py
@@ -76,7 +76,7 @@ if __name__ == "__main__":
Inputs.Marginals[1].Name = 'x2'
Inputs.Marginals[1].DistType = 'unif'
Inputs.Marginals[1].Parameters = [-5, 5]
-
+
#=====================================================
#====== POLYNOMIAL CHAOS EXPANSION METAMODELS ======
#=====================================================
@@ -86,8 +86,8 @@ if __name__ == "__main__":
# The degree with the lowest Leave-One-Out cross-validation (LOO)
# error (or the highest score=1-LOO)estimator is chosen as the final
# metamodel.
- MetaModelOpts.MinPceDegree = 2 #2
- MetaModelOpts.MaxPceDegree = 8 #8
+ MetaModelOpts.MinPceDegree = 4 #2
+ MetaModelOpts.MaxPceDegree = 10 #8
# q-quasi-norm 0<q<1 (default=1)
MetaModelOpts.q = 1 #0.75
@@ -97,7 +97,7 @@ if __name__ == "__main__":
# 1)AaPCE: Adaptive aPCE 2)BRR: Bayesian Ridge Regression
# 3)LARS: Least angle regression 4)ARD: Bayesian ARD Regression (Non-degree Adaptive)
- MetaModelOpts.RegMethod = 'ARD'
+ MetaModelOpts.RegMethod = 'BRR'
# Print summary of the regression results
#MetaModelOpts.DisplayFlag = True
@@ -143,21 +143,21 @@ if __name__ == "__main__":
MetaModelOpts.ExpDesign.MaxFunItr = 200
# Use when 'Voronoi' or 'MC' or 'LHS' chosen
- MetaModelOpts.ExpDesign.NCandidate = 5
+ MetaModelOpts.ExpDesign.NCandidate = 3
MetaModelOpts.ExpDesign.NrofCandGroups = 4
# -------- Exploitation ------
# 1)'BayesOptDesign' 2)'VarOptDesign' 3)'alphabetic' 4)'Space-filling'
- MetaModelOpts.ExpDesign.ExploitMethod = 'VarOptDesign'
+ MetaModelOpts.ExpDesign.ExploitMethod = 'BayesOptDesign'
# BayesOptDesign -> when data is available
# 1)DKL (Kullback-Leibler Divergence) 2)DPP (D-Posterior-percision)
# 3)APP (A-Posterior-percision)
- #MetaModelOpts.ExpDesign.UtilityFunction = 'DKL' #['DKL', 'DPP']
+ MetaModelOpts.ExpDesign.UtilityFunction = 'DKL' #['DKL', 'DPP']
# VarBasedOptDesign -> when data is not available
# Only with Vornoi >>> 1)Entropy 2)EIGF, 3)ALM, 4)LOOCV
- MetaModelOpts.ExpDesign.UtilityFunction = 'Entropy'#['EIGF', 'Entropy', 'LOOCV']
+ #MetaModelOpts.ExpDesign.UtilityFunction = 'Entropy'#['EIGF', 'Entropy', 'LOOCV']
# alphabetic
# 1)D-Opt (D-Optimality) 2)A-Opt (A-Optimality)
--
GitLab