.. only:: html
.. note::
:class: sphx-glr-download-link-note
Click :ref:`here ` to download the full example code or to run this example in your browser via Binder
.. rst-class:: sphx-glr-example-title
.. _sphx_glr_auto_examples_ask-and-tell.py:
=======================
Async optimization Loop
=======================
Bayesian optimization is used to tune parameters for walking robots or other
experiments that are not a simple (expensive) function call.
Tim Head, February 2017.
Reformatted by Holger Nahrstaedt 2020
.. currentmodule:: skopt
They often follow a pattern a bit like this:
1. ask for a new set of parameters
2. walk to the experiment and program in the new parameters
3. observe the outcome of running the experiment
4. walk back to your laptop and tell the optimizer about the outcome
5. go to step 1
A setup like this is difficult to implement with the ***_minimize()** function
interface. This is why **scikit-optimize** has a ask-and-tell interface that
you can use when you want to control the execution of the optimization loop.
This notebook demonstrates how to use the ask and tell interface.
.. code-block:: default
print(__doc__)
import numpy as np
np.random.seed(1234)
import matplotlib.pyplot as plt
from skopt.plots import plot_gaussian_process
The Setup
---------
We will use a simple 1D problem to illustrate the API. This is a little bit
artificial as you normally would not use the ask-and-tell interface if you
had a function you can call to evaluate the objective.
.. code-block:: default
from skopt.learning import ExtraTreesRegressor
from skopt import Optimizer
noise_level = 0.1
Our 1D toy problem, this is the function we are trying to
minimize
.. code-block:: default
def objective(x, noise_level=noise_level):
return np.sin(5 * x[0]) * (1 - np.tanh(x[0] ** 2))\
+ np.random.randn() * noise_level
def objective_wo_noise(x, noise_level=0):
return objective(x, noise_level=0)
Here a quick plot to visualize what the function looks like:
.. code-block:: default
# Plot f(x) + contours
plt.set_cmap("viridis")
x = np.linspace(-2, 2, 400).reshape(-1, 1)
fx = np.array([objective(x_i, noise_level=0.0) for x_i in x])
plt.plot(x, fx, "r--", label="True (unknown)")
plt.fill(np.concatenate([x, x[::-1]]),
np.concatenate(([fx_i - 1.9600 * noise_level for fx_i in fx],
[fx_i + 1.9600 * noise_level for fx_i in fx[::-1]])),
alpha=.2, fc="r", ec="None")
plt.legend()
plt.grid()
plt.show()
.. image:: /auto_examples/images/sphx_glr_ask-and-tell_001.png
:alt: ask and tell
:class: sphx-glr-single-img
Now we setup the :class:`Optimizer` class. The arguments follow the meaning and
naming of the ***_minimize()** functions. An important difference is that
you do not pass the objective function to the optimizer.
.. code-block:: default
opt = Optimizer([(-2.0, 2.0)], "GP", acq_func="EI",
acq_optimizer="sampling",
initial_point_generator="lhs")
# To obtain a suggestion for the point at which to evaluate the objective
# you call the ask() method of opt:
next_x = opt.ask()
print(next_x)
.. rst-class:: sphx-glr-script-out
Out:
.. code-block:: none
[-0.7315058981975282]
In a real world use case you would probably go away and use this
parameter in your experiment and come back a while later with the
result. In this example we can simply evaluate the objective function
and report the value back to the optimizer:
.. code-block:: default
f_val = objective(next_x)
opt.tell(next_x, f_val)
.. rst-class:: sphx-glr-script-out
Out:
.. code-block:: none
fun: 0.2071864923643295
func_vals: array([0.20718649])
models: []
random_state: RandomState(MT19937) at 0x7F46AE13E640
space: Space([Real(low=-2.0, high=2.0, prior='uniform', transform='normalize')])
specs: {'args': {'self': , 'dimensions': [(-2.0, 2.0)], 'base_estimator': 'GP', 'n_random_starts': None, 'n_initial_points': 10, 'initial_point_generator': 'lhs', 'n_jobs': 1, 'acq_func': 'EI', 'acq_optimizer': 'sampling', 'random_state': None, 'model_queue_size': None, 'acq_func_kwargs': None, 'acq_optimizer_kwargs': None}, 'function': 'Optimizer'}
x: [-0.7315058981975282]
x_iters: [[-0.7315058981975282]]
Like ***_minimize()** the first few points are suggestions from
the initial point generator as there
is no data yet with which to fit a surrogate model.
.. code-block:: default
for i in range(9):
next_x = opt.ask()
f_val = objective(next_x)
res = opt.tell(next_x, f_val)
We can now plot the random suggestions and the first model that has been
fit:
.. code-block:: default
_ = plot_gaussian_process(res, objective=objective_wo_noise,
noise_level=noise_level,
show_next_point=False,
show_acq_func=True)
plt.show()
.. image:: /auto_examples/images/sphx_glr_ask-and-tell_002.png
:alt: x* = -0.3201, f(x*) = -0.9482
:class: sphx-glr-single-img
Let us sample a few more points and plot the optimizer again:
.. code-block:: default
for i in range(10):
next_x = opt.ask()
f_val = objective(next_x)
res = opt.tell(next_x, f_val)
_ = plot_gaussian_process(res, objective=objective_wo_noise,
noise_level=noise_level,
show_next_point=True,
show_acq_func=True)
plt.show()
.. image:: /auto_examples/images/sphx_glr_ask-and-tell_003.png
:alt: x* = -0.3201, f(x*) = -0.9482
:class: sphx-glr-single-img
By using the :class:`Optimizer` class directly you get control over the
optimization loop.
You can also pickle your :class:`Optimizer` instance if you want to end the
process running it and resume it later. This is handy if your experiment
takes a very long time and you want to shutdown your computer in the
meantime:
.. code-block:: default
import pickle
with open('my-optimizer.pkl', 'wb') as f:
pickle.dump(opt, f)
with open('my-optimizer.pkl', 'rb') as f:
opt_restored = pickle.load(f)
.. rst-class:: sphx-glr-timing
**Total running time of the script:** ( 0 minutes 2.768 seconds)
**Estimated memory usage:** 14 MB
.. _sphx_glr_download_auto_examples_ask-and-tell.py:
.. only :: html
.. container:: sphx-glr-footer
:class: sphx-glr-footer-example
.. container:: binder-badge
.. image:: /../../miniconda/envs/testenv/lib/python3.8/site-packages/sphinx_gallery/_static/binder_badge_logo.svg
:target: https://mybinder.org/v2/gh/scikit-optimize/scikit-optimize/master?urlpath=lab/tree/notebooks/auto_examples/ask-and-tell.ipynb
:width: 150 px
.. container:: sphx-glr-download sphx-glr-download-python
:download:`Download Python source code: ask-and-tell.py `
.. container:: sphx-glr-download sphx-glr-download-jupyter
:download:`Download Jupyter notebook: ask-and-tell.ipynb `
.. only:: html
.. rst-class:: sphx-glr-signature
`Gallery generated by Sphinx-Gallery `_