The Developer API is suitable when the user wants maximal customization of the optimization loop. This tutorial demonstrates optimization of a Hartmann6 function using the SimpleExperiment construct, which we use for synchronous experiments, where trials can be evaluated right away.
For more details on the different Ax constructs, see the "Building Blocks of Ax" tutorial.
import numpy as np
from ax import (
ComparisonOp,
ParameterType,
RangeParameter,
SearchSpace,
SimpleExperiment,
OutcomeConstraint,
)
from ax.metrics.l2norm import L2NormMetric
from ax.modelbridge.registry import Models
from ax.plot.contour import plot_contour
from ax.plot.trace import optimization_trace_single_method
from ax.utils.measurement.synthetic_functions import hartmann6
from ax.utils.notebook.plotting import render, init_notebook_plotting
init_notebook_plotting()
[INFO 07-17 16:19:19] ipy_plotting: Injecting Plotly library into cell. Do not overwrite or delete cell.
First, we define an evaluation function that is able to compute all the metrics needed for this experiment. This function needs to accept a set of parameter values and can also accept a weight. It should produce a dictionary of metric names to tuples of mean and standard error for those metrics. Note that when using Experiment (instead of SimpleExperiment), it's possible to deploy trials and fetch their evaluation results asynchronously; more on that in the "Building Blocks of Ax" tutorial.
def hartmann_evaluation_function(
parameterization, # Mapping of parameter names to values of those parameters.
weight=None, # Optional weight argument.
):
x = np.array([parameterization.get(f"x{i}") for i in range(6)])
# In our case, standard error is 0, since we are computing a synthetic function.
return {"hartmann6": (hartmann6(x), 0.0), "l2norm": (np.sqrt((x ** 2).sum()), 0.0)}
If there is only one metric in the experiment – the objective – then evaluation function can return a single tuple of mean and SEM, in which case Ax will assume that evaluation corresponds to the objective. It can also return only the mean as a float, in which case Ax will treat SEM as unknown and use a model that can infer it. For more details on evaluation function, refer to the "Trial Evaluation" section in the docs.
Second, we define a search space, which defines the type and allowed range for the parameters.
hartmann_search_space = SearchSpace(
parameters=[
RangeParameter(
name=f"x{i}", parameter_type=ParameterType.FLOAT, lower=0.0, upper=1.0
)
for i in range(6)
]
)
Third, we make a SimpleExperiment. In addition to the search space and evaluation function, here we define the objective_name and outcome_constraints.
When doing the optimization, we will find points that minimize the objective while obeying the constraints (which in this case means l2norm < 1.25).
exp = SimpleExperiment(
name="test_hartmann",
search_space=hartmann_search_space,
evaluation_function=hartmann_evaluation_function,
objective_name="hartmann6",
minimize=True,
outcome_constraints=[
OutcomeConstraint(
metric=L2NormMetric(
name="l2norm", param_names=[f"x{i}" for i in range(6)], noise_sd=0.2
),
op=ComparisonOp.LEQ,
bound=1.25,
relative=False,
)
],
)
Run the optimization using the settings defined on the experiment. We will create 5 random sobol points for exploration followed by 15 points generated using the GPEI optimizer.
print(f"Running Sobol initialization trials...")
sobol = Models.SOBOL(exp.search_space)
for i in range(5):
exp.new_trial(generator_run=sobol.gen(1))
for i in range(25):
print(f"Running GP+EI optimization trial {i+1}/15...")
# Reinitialize GP+EI model at each step with updated data.
gpei = Models.BOTORCH(experiment=exp, data=exp.eval())
batch = exp.new_trial(generator_run=gpei.gen(1))
print("Done!")
Running Sobol initialization trials... Running GP+EI optimization trial 1/15... Running GP+EI optimization trial 2/15... Running GP+EI optimization trial 3/15... Running GP+EI optimization trial 4/15... Running GP+EI optimization trial 5/15... Running GP+EI optimization trial 6/15... Running GP+EI optimization trial 7/15... Running GP+EI optimization trial 8/15... Running GP+EI optimization trial 9/15... Running GP+EI optimization trial 10/15... Running GP+EI optimization trial 11/15... Running GP+EI optimization trial 12/15... Running GP+EI optimization trial 13/15... Running GP+EI optimization trial 14/15... Running GP+EI optimization trial 15/15... Running GP+EI optimization trial 16/15... Running GP+EI optimization trial 17/15... Running GP+EI optimization trial 18/15... Running GP+EI optimization trial 19/15... Running GP+EI optimization trial 20/15... Running GP+EI optimization trial 21/15... Running GP+EI optimization trial 22/15... Running GP+EI optimization trial 23/15... Running GP+EI optimization trial 24/15... Running GP+EI optimization trial 25/15... Done!
Now we can inspect the SimpleExperiment's data by calling eval(), which retrieves evaluation data for all batches of the experiment.
We can also use the eval_trial function to get evaluation data for a specific trial in the experiment, like so:
trial_data = exp.eval_trial(exp.trials[1])
trial_data.df
| arm_name | metric_name | mean | sem | trial_index | |
|---|---|---|---|---|---|
| 0 | 1_0 | hartmann6 | -0.387304 | 0.0 | 1 |
| 1 | 1_0 | l2norm | 1.108547 | 0.0 | 1 |
Now we can plot the results of our optimization:
# `plot_single_method` expects a 2-d array of means, because it expects to average means from multiple
# optimization runs, so we wrap out best objectives array in another array.
objective_means = np.array([[trial.objective_mean for trial in exp.trials.values()]])
best_objective_plot = optimization_trace_single_method(
y=np.minimum.accumulate(objective_means, axis=1),
optimum=-3.32237, # Known minimum objective for Hartmann6 function.
)
render(best_objective_plot)
Total runtime of script: 1 minutes, 27.96 seconds.