tfp.experimental.bayesopt.acquisition.MCMCReducer

Acquisition function for reducing over batch dimensions.

Inherits From: AcquisitionFunction

tfp.experimental.bayesopt.acquisition.MCMCReducer(
    predictive_distribution,
    observations,
    seed=None,
    acquisition_class=None,
    reduce_dims=None,
    **acquisition_kwargs
)

MCMCReducer evaluates a base acquisition function and takes the mean of the function values over the dimensions indicated by reduce_dims. MCMCReducer is useful for marginalizing over an MCMC sample of GP kernel hyperparameters, for example.

Examples

Build and evaluate an acquisition function that computes Gaussian Process Expected Improvement and then marginalizes over the leftmost batch dimension.

import numpy as np
import tensorflow_probability as tfp

tfd = tfp.distributions
tfpk = tfp.math.psd_kernels
tfp_acq = tfp.experimental.bayesopt.acquisition

# Sample 10 20-dimensional index points and associated observations.
index_points = np.random.uniform(size=[10, 20])
observations = np.random.uniform(size=[10])

# The kernel and GP have batch shape [32], representing a sample of
# hyperparameters that we want to marginalize over.
kernel_amplitudes = np.random.uniform(size=[32])

# Build a (batched) Gaussian Process regression model.
dist = tfd.GaussianProcessRegressionModel(
    kernel=tfpk.MaternFiveHalves(amplitude=kernel_amplitudes),
    observation_index_points=index_points,
    observations=observations)

# Define an `MCMCReducer` with GP Expected Improvement.
mcmc_ei = tfp_acq.MCMCReducer(
    predictive_distribution=dist,
    observations=observations,
    acquisition_class=GaussianProcessExpectedImprovement,
    reduce_dims=0)

# Evaluate the acquisition function at a new set of index points,
# marginalizing over the hyperparameter batch.
pred_index_points = np.random.uniform(size=[6, 20])
acq_fn_vals = mcmc_ei(pred_index_points)  # Has shape [6].

Args
predictive_distribution	tfd.Distribution-like, the distribution over observations at a set of index points.
observations	Float Tensor of observations.
seed	PRNG seed; see tfp.random.sanitize_seed for details.
acquisition_class	AcquisitionFunction-like callable.
reduce_dims	Axis of the acquisition function value array over which to marginalize.
**acquisition_kwargs	Kwargs passed to acquisition_class.

Attributes
acquisition
is_parallel	Python bool indicating whether the acquisition function is parallel. Parallel (batched) acquisition functions evaluate batches of points rather than single points.
observations	Float Tensor of observations.
predictive_distribution	The distribution over observations at a set of index points.
reduce_dims
seed	PRNG seed.

Methods

__call__

View source

__call__(
    **kwargs
)

Call self as a function.