Merge branch 'master' into gradients-for-covariance-gpy

.github/workflows/tests.yml

@@ -12,7 +12,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: [3.7]
+        python-version: [3.7, 3.8]
 
     steps:
     - uses: actions/checkout@v2
@@ -30,12 +30,12 @@ jobs:
         # work around issues with GPy setting matplotlib backend
         echo 'backend: Agg' > matplotlibrc
         pip install .
-    # - name: Lint with flake8
-    #   run: |
-    #     # stop the build if there are Python syntax errors or undefined names
-    #     flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
-    #     # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
-    #     flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics
+    - name: Lint with flake8
+      run: |
+        # stop the build if there are Python syntax errors or undefined names
+        flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+        # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+        flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics
     - name: Unit tests
       run: |
         python -m pytest

CHANGELOG.md

@@ -1,6 +1,16 @@
 # Changelog
 All notable changes to Emukit will be documented in this file.
 
+## [0.4.8]
+- Added sobol initial design
+- BanditParameter
+- Boolean operations for stopping conditions
+- Preferential Bayesian optimization example
+- MUMBO acquisition function
+- Revised dependecies' versions requirements
+- Bug fixes
+- Doc fixes
+
 ## [0.4.7]
 - Added simple GP model for examples
 - Bayesian optimization with unknown constraints

emukit/__version__.py

@@ -11,4 +11,4 @@
 #   express or implied. See the License for the specific language governing
 #   permissions and limitations under the License.
 # ==============================================================================
-__version__ = '0.4.7'  # noqa
+__version__ = '0.4.8'  # noqa

emukit/examples/dynamic_negative_lower_confidence_bound/__init__.py

@@ -0,0 +1 @@
+from .dnlcb import DynamicNegativeLowerConfidenceBound

emukit/examples/dynamic_negative_lower_confidence_bound/dnlcb.py

@@ -0,0 +1,31 @@
+from emukit.bayesian_optimization.acquisitions import NegativeLowerConfidenceBound
+from emukit.core.interfaces import IModel, IDifferentiable
+import numpy as np
+from typing import Union
+
+class DynamicNegativeLowerConfidenceBound(NegativeLowerConfidenceBound):
+
+    def __init__(self, model: Union[IModel, IDifferentiable], input_space_size: int, delta: float) -> None:
+        """
+        Dynamic extension of the LCB acquisition. The beta coefficient is updated at each iteration, based on the explorativeness parameter delta which is inversely
+        proportional to beta itself - the higher the delta the less explorative the selection will be.
+        Please consider that regret bounds based on the dynamic exploration coefficient only hold for selected kernel classes exhibiting boundedness and smoothness.
+        See the base class for paper references.
+        This class may also be taken as a reference for acquisition functions that dynamically update their parameters thanks to the update_parameters() hook; the implicit assumption is that this method is invoked once per iteration (it is no big deal if this is the case for a constant number of times per iteration; should it be more then we are increasing the beta too fast).
+        :param model: The underlying model that provides the predictive mean and variance for the given test points
+        :param input_space_size: the size of the finite D grid on which the function is evaluated
+        :param delta: the exploration parameter determining the beta exploration coefficient; delta must be in (0, 1) and it is inversely related to beta
+        """
+        assert input_space_size > 0, "Invalid dimension provided"
+        assert 0 < delta < 1, "Delta must be in (0, 1)"
+        super().__init__(model)
+        self.input_space_size = input_space_size
+        self.delta = delta
+        self.iteration = 0
+
+    def optimal_beta_selection(self) -> float:
+        return 2 * np.log(self.input_space_size * (self.iteration ** 2) * (np.pi ** 2) / (6 * self.delta))
+
+    def update_parameters(self) -> None:
+        self.iteration += 1
+        self.beta = self.optimal_beta_selection()

emukit/model_wrappers/gpy_model_wrappers.py

@@ -147,7 +147,14 @@ def generate_hyperparameters_samples(self, n_samples=20, n_burnin=100, subsample
 
         """
         self.model.optimize(max_iters=self.n_restarts)
-        self.model.param_array[:] = self.model.param_array * (1. + np.random.randn(self.model.param_array.size) * 0.01)
+        # Add jitter to all unfixed parameters. After optimizing the hyperparameters, the gradient of the
+        # posterior probability of the parameters wrt. the parameters will be close to 0.0, which is a poor
+        # initialization for HMC
+        unfixed_params = [param for param in self.model.flattened_parameters if not param.is_fixed]
+        for param in unfixed_params:
+            # Add jitter by multiplying with log-normal noise with mean 1 and standard deviation 0.01 
+            # This ensures the sign of the parameter remains the same
+            param *= np.random.lognormal(np.log(1. / np.sqrt(1.0001)), np.sqrt(np.log(1.0001)), size=param.size)
         hmc = GPy.inference.mcmc.HMC(self.model, stepsize=step_size)
         samples = hmc.sample(num_samples=n_burnin + n_samples * subsample_interval, hmc_iters=leapfrog_steps)
         hmc_samples = samples[n_burnin::subsample_interval]

notebooks/index.ipynb

@@ -54,7 +54,8 @@
     "* [How to perform Bayesian optimization with non-linear constraints](Emukit-tutorial-constrained-optimization.ipynb)\n",
     "* [Bayesian optimization integrating the hyper-parameters of the model](Emukit-tutorial-bayesian-optimization-integrating-model-hyperparameters.ipynb)\n",
     "* [How to use custom model](Emukit-tutorial-custom-model.ipynb)\n",
-    "* [How to select neural network hyperparameters](Emukit-tutorial-select-neural-net-hyperparameters.ipynb)"
+    "* [How to select neural network hyperparameters](Emukit-tutorial-select-neural-net-hyperparameters.ipynb)\n",
+    "* [How to parallelize external objective function evaluations in Bayesian optimization](Emukit-tutorial-parallel-eval-of-obj-fun.ipynb)"
    ]
   },
   {