UPDATE: docs for ts

docs/conf.py

@@ -24,7 +24,7 @@
 author = 'Jungtaek Kim and Seungjin Choi'
 
 # The short X.Y version
-version = '0.4.0'
+version = '0.4.1'
 # The full version, including alpha/beta/rc tags
 release = '{} alpha'.format(version)
 

docs/example/ts_gp_prior.rst

@@ -0,0 +1,136 @@
+Optimizing a sampled function via Thompson sampling
+===================================================
+
+This example is to optimize a function sampled from a Gaussian process prior via Thompson sampling.
+First of all, import the packages we need and **bayeso**.
+
+.. code-block:: python
+
+    import numpy as np
+
+    from bayeso import gp
+    from bayeso import covariance
+    from bayeso.utils import utils_covariance
+    from bayeso.utils import utils_plotting
+
+Declare some parameters to control this example, including zero-mean prior, and compute a covariance matrix.
+
+.. code-block:: python
+
+    num_points = 1000
+    str_cov = 'se'
+    int_init = 1
+    int_iter = 50
+    int_ts = 100
+
+    list_Y_min = []
+
+    X = np.expand_dims(np.linspace(-5, 5, num_points), axis=1)
+    mu = np.zeros(num_points)
+    hyps = utils_covariance.get_hyps(str_cov, 1)
+    Sigma = covariance.cov_main(str_cov, X, X, hyps, True)
+
+Optimize a function sampled from a Gaussian process prior.
+At each iteration, we sample a query point that outputs the mininum value of the function sampled from a Gaussian process posterior.
+
+.. code-block:: python
+
+    for ind_ts in range(0, int_ts):
+        print('TS:', ind_ts + 1, 'iteration')
+        Y = gp.sample_functions(mu, Sigma, num_samples=1)[0]
+
+        ind_init = np.argmin(Y)
+        bx_min = X[ind_init]
+        y_min = Y[ind_init]
+    
+        ind_random = np.random.choice(num_points)
+
+        X_ = np.expand_dims(X[ind_random], axis=0)
+        Y_ = np.expand_dims(np.expand_dims(Y[ind_random], axis=0), axis=1)
+
+        for ind_iter in range(0, int_iter):
+            print(ind_iter + 1, 'iteration')
+
+            mu_, sigma_, Sigma_ = gp.predict_optimized(X_, Y_, X, str_cov=str_cov)
+            ind_ = np.argmin(gp.sample_functions(np.squeeze(mu_, axis=1), Sigma_, num_samples=1)[0])
+
+            X_ = np.concatenate([X_, [X[ind_]]], axis=0)
+            Y_ = np.concatenate([Y_, [[Y[ind_]]]], axis=0)
+        
+        list_Y_min.append(Y_ - y_min)
+
+    Ys = np.array(list_Y_min)
+    Ys = np.squeeze(Ys, axis=2)
+    print(Ys.shape)
+
+Plot the result obtained from the code block above.
+
+.. code-block:: python
+
+    utils_plotting.plot_minimum(np.array([Ys]), ['TS'], 1, True,
+                                is_tex=True, range_shade=1.0,
+                                str_x_axis=r'\textrm{Iteration}',
+                                str_y_axis=r'\textrm{Minimum regret}')
+
+.. image:: ../_static/examples/ts_gp_prior.*
+    :width: 320
+    :align: center
+    :alt: ts_gp_prior
+
+Full code:
+
+.. code-block:: python
+
+    import numpy as np
+
+    from bayeso import gp
+    from bayeso import covariance
+    from bayeso.utils import utils_covariance
+    from bayeso.utils import utils_plotting
+
+    num_points = 1000
+    str_cov = 'se'
+    int_init = 1
+    int_iter = 50
+    int_ts = 100
+
+    list_Y_min = []
+
+    X = np.expand_dims(np.linspace(-5, 5, num_points), axis=1)
+    mu = np.zeros(num_points)
+    hyps = utils_covariance.get_hyps(str_cov, 1)
+    Sigma = covariance.cov_main(str_cov, X, X, hyps, True)
+
+    for ind_ts in range(0, int_ts):
+        print('TS:', ind_ts + 1, 'iteration')
+        Y = gp.sample_functions(mu, Sigma, num_samples=1)[0]
+
+        ind_init = np.argmin(Y)
+        bx_min = X[ind_init]
+        y_min = Y[ind_init]
+    
+        ind_random = np.random.choice(num_points)
+
+        X_ = np.expand_dims(X[ind_random], axis=0)
+        Y_ = np.expand_dims(np.expand_dims(Y[ind_random], axis=0), axis=1)
+
+        for ind_iter in range(0, int_iter):
+            print(ind_iter + 1, 'iteration')
+
+            mu_, sigma_, Sigma_ = gp.predict_optimized(X_, Y_, X, str_cov=str_cov)
+            ind_ = np.argmin(gp.sample_functions(np.squeeze(mu_, axis=1), Sigma_, num_samples=1)[0])
+
+            X_ = np.concatenate([X_, [X[ind_]]], axis=0)
+            Y_ = np.concatenate([Y_, [[Y[ind_]]]], axis=0)
+        
+        list_Y_min.append(Y_ - y_min)
+
+    Ys = np.array(list_Y_min)
+    Ys = np.squeeze(Ys, axis=2)
+    print(Ys.shape)
+
+    utils_plotting.plot_minimum(np.array([Ys]), ['TS'], 1, True,
+                                is_tex=True, range_shade=1.0,
+                                str_x_axis=r'\textrm{Iteration}',
+                                str_y_axis=r'\textrm{Minimum regret}')
+

docs/index.rst

@@ -33,6 +33,7 @@ The code can be found in `our GitHub repository <https://github.com/jungtaekkim/
    :caption: Example:
 
    example/gp
+   example/ts_gp_prior
    example/branin
    example/hpo
 

examples/99_notebooks/example_ts_gp_prior.ipynb

@@ -9,8 +9,6 @@
     "%matplotlib inline\n",
     "\n",
     "import numpy as np\n",
-    "import os\n",
-    "import matplotlib.pyplot as plt\n",
     "\n",
     "from bayeso import gp\n",
     "from bayeso import covariance\n",
@@ -80,7 +78,7 @@
    "outputs": [],
    "source": [
     "utils_plotting.plot_minimum(np.array([Ys]), ['TS'], 1, True,\n",
-    "                            is_tex=True,\n",
+    "                            is_tex=True, range_shade=1.0,\n",
     "                            str_x_axis=r'\\textrm{Iteration}',\n",
     "                            str_y_axis=r'\\textrm{Minimum regret}')"
    ]