Added updated objective function for hyper parameter optimization, se…

…ems to be working for sigma values, but not for kernel parameters

examples/GaussianProcess/User2DGP.py

@@ -57,7 +57,7 @@ def f_sq(x, data=None):
     #gp = gr.PreferenceGP(gr.linear_kern(0.3, 0.1, 0.0))
     gp = gr.PreferenceGP(gr.RBF_kern(1.0, 1.0), pareto_pairs=True, \
                         use_hyper_optimization=False, normalize_positive=True)
-    gp.add_prior(bounds=np.array(bounds))
+    #gp.add_prior(bounds=np.array(bounds), num_pts=100)
 
 
 
@@ -78,7 +78,7 @@ def f_sq(x, data=None):
 
         gp.add(train_X[selected_idx], pairs)
 
-    gp.optimize(optimize_hyperparameter=False)
+    gp.optimize(optimize_hyperparameter=True)
 
     x = np.linspace(bounds[0][0], bounds[0][1], num_side)
     y = np.linspace(bounds[1][0], bounds[1][1], num_side)

examples/GaussianProcess/UserGP.py

@@ -39,18 +39,20 @@ def f_sin(x, data=None):
             gr.generate_fake_pairs(X_train, f_sin, 4)
 
 
-    gp = gr.PreferenceGP(gr.RBF_kern(0.5, 0.7))
+    #gp = gr.PreferenceGP(gr.RBF_kern(0.5, 0.7))
+    gp = gr.PreferenceGP(gr.RBF_kern(0.5, 0.2))
     #gp = gr.PreferenceGP(gr.periodic_kern(1.2,0.3,5))
     #gp = gr.PreferenceGP(gr.linear_kern(0.2, 0.2, 0.2))
     #gp = gr.PreferenceGP(gr.RBF_kern(0.2,1)+gr.periodic_kern(1,0.2,0)+gr.linear_kern(0.2,0.1,0.3))
     #gp = gr.PreferenceGP(gr.RBF_kern(0.1,1)*gr.linear_kern(0.3,0.2,0.3))
 
     gp.add(X_train, pairs)
 
-    gp.optimize(optimize_hyperparameter=False)
+    gp.optimize(optimize_hyperparameter=True)
     print('gp.calc_ll()')
     print(gp.calc_ll())
 
+    print('Simga = '+str(gp.probits[0].sigma))
 
     X = np.arange(-0.5, 8, 0.1)
     mu, sigma = gp.predict(X)

src/rdml_graph/gaussian_process/PreferenceGP.py

@@ -67,7 +67,7 @@ def __init__(self, cov_func, normalize_gp=True, pareto_pairs=False, \
         super(PreferenceGP, self).__init__(cov_func, mat_inv, active_learner=active_learner)
         PreferenceModel.__init__(self, pareto_pairs, other_probits)
 
-        self.lambda_gp = 0.1
+        self.lambda_gp = 0.9
 
         self.normalize_gp = normalize_gp
 
@@ -76,7 +76,7 @@ def __init__(self, cov_func, normalize_gp=True, pareto_pairs=False, \
 
         # sigma on the likelihood function.
         #self.sigma_L = 1.0
-        self.probits = [PreferenceProbit(sigma = 1.0)]
+        self.probits = [PreferenceProbit(sigma = 0.05)]
 
 
 
@@ -88,7 +88,7 @@ def __init__(self, cov_func, normalize_gp=True, pareto_pairs=False, \
 
 
 
-        self.delta_f = 0.002 # set the convergence to stop
+        self.delta_f = 0.0002 # set the convergence to stop
         self.maxloops = 100
 
 
@@ -144,14 +144,15 @@ def calc_ll_param(self, hyperparameters, X_train, y_train):
 
         self.probits[self.probit_idxs['relative_discrete']].set_sigma(sigma_L)
         self.cov_func.set_param(hyperparameters[1:])
-        K = self.cov_func.cov(X_train, X_train)
+        # K = self.cov_func.cov(X_train, X_train)
 
-        W, dpy_df, logpYF = self.derivatives(y_train, self.F)
+        # W, dpy_df, logpYF = self.derivatives(y_train, self.F)
 
-        Kinv = self.invert_function(K)
-        term2 = 0.5 * np.matmul(np.matmul(np.transpose(self.F), Kinv), self.F)
+        # Kinv = self.invert_function(K)
+        # term2 = 0.5 * np.matmul(np.matmul(np.transpose(self.F), Kinv), self.F)
 
-        term3 = 0.5 * np.log(np.linalg.det(np.identity(len(K)) + np.matmul(K, W)))
+        # term3 = 0.5 * np.log(np.linalg.det(np.identity(len(K)) + np.matmul(K, W)))
+        return self.loss_F(self.F)
 
         #return logpYF
         return logpYF - term2 - term3
@@ -254,6 +255,7 @@ def findMode(self, x_train, y_train):
 
             # check for convergence
             f_err = np.linalg.norm((F_new - F), ord=np.inf)
+            print("\tf_err="+str(f_err))
             F = F_new
 
 
@@ -262,11 +264,16 @@ def findMode(self, x_train, y_train):
                 print('WARNING: maximum loops in findMode exceeded. Returning current solution')
                 break
 
+        print('Optimization ran for: '+str(n_loops))
+
+
         self.F = F
         # calculate W with final F
         self.W, self.grad_ll, self.log_likelihood = \
                                         self.derivatives(y_train, self.F)
 
+        print('Best solution, training data likelyhood = '+str(self.log_likelihood))
+
 
     ## optimize
     # Runs the optimization step required by the user preference GP.

src/rdml_graph/gaussian_process/PreferenceModel.py

@@ -237,6 +237,9 @@ def add(self, X, y, type='relative_discrete', training_sigma=0):
                     self.y_train[self.probit_idxs['relative_discrete']] = \
                         np.append(self.y_train[self.probit_idxs['relative_discrete']], \
                                     np.array(pairs), axis=0)
+                else:
+                    import pdb
+                    pdb.set_trace()
         # end if for pareto_pairs