From 4424fc0eac546464b2731613c3fbb61568b4f114 Mon Sep 17 00:00:00 2001
From: Hu Zhao <zhao@mbd.rwth-aachen.de>
Date: Sat, 22 Apr 2023 05:07:49 +0200
Subject: [PATCH] docs: add example of active learning
---
.../inference/plot_active_learning.py | 127 ++++++++++++++++++
1 file changed, 127 insertions(+)
create mode 100644 docs/examples/inference/plot_active_learning.py
diff --git a/docs/examples/inference/plot_active_learning.py b/docs/examples/inference/plot_active_learning.py
new file mode 100644
index 0000000..5f341c0
--- /dev/null
+++ b/docs/examples/inference/plot_active_learning.py
@@ -0,0 +1,127 @@
+"""
+
+Active learning
+===============
+"""
+
+# %% md
+#
+# This example shows how to use the :class:`.ActiveLearning` class to iteratively
+# build a Gaussian process emulator for an unnormalized posterior involving a
+# simulator. It should be noted that this example is only for illustration
+# purpose, rather than a real case. For simplicity, required arguments for
+# :class:`.ActiveLearning` (including `simulator`, `likelihood`, `data`, etc.) are purely
+# made up. For a realistic case of active learning, one can refer to :cite:t:`Zhao2022`.
+#
+# First, we define the simulator, prior distribution of its variable parameters,
+# observed data, and likelihood function. They basically define the Bayesian
+# inference problem.
+
+import numpy as np
+
+ndim = 2 # dimension of variable parameters of the simulator
+bounds = bounds = np.array([[-5,5],[-5,5]]) # bounds of variable parameters
+data = np.array([1,0])
+
+def simulator(x1, x2):
+ """Simulator y=f(x)."""
+ y1, y2 = x1, x2
+ return np.array([y1, y2])
+
+def prior(x):
+ """Uniform prior."""
+ return 1/(10*10)
+
+def likelihood(y, data):
+ """Likelihood function L(y,data)."""
+ return np.exp(-(y[0]-data[0])**2/100 - (y[0]**2-y[1]-data[1])**2)
+
+# %% md
+#
+# Imagine that the simulator is a complex solver. It is not computationally feasible
+# to compute the posterior distribution of the variable parameters using grid estimation
+# or Metropolis Hastings estimation. This is because they require evaluating the likelihood
+# many times which essentially leads to many evaluations of the simulator. Therefore, we
+# resort to use active learning to build a Gaussian process emulator for the unnormalized
+# posterior (prior times likelihood) based on a small number of evaluations of the simulator.
+# The the posterior can be estimated using the emulator.
+#
+# To do so, we need to pass arguments to following parameters of the :class:`.ActiveLearning` class:
+#
+# - run_sim_obj : instance of class :class:`.RunSimulator`. It carries information on how
+# to run the simulator.
+# - lhs_sampler : instance of class :class:`.LHS`. It is used to draw initial samples to run
+# simulations in order to train an inital Gaussian process emulator.
+# - scalar_gasp : instance of class :class:`.ScalarGaSP`. It sets up the emulator structure.
+#
+
+from psimpy.simulator import RunSimulator
+from psimpy.sampler import LHS
+from psimpy.emulator import ScalarGaSP
+
+run_simulator = RunSimulator(simulator, var_inp_parameter=['x1','x2'])
+lhs_sampler = LHS(ndim=ndim, bounds=bounds, seed=1)
+scalar_gasp = ScalarGaSP(ndim=ndim)
+
+# %%md
+#
+# Next, we create an object of the :class:`.ActiveLearning` class by
+
+from psimpy.inference import ActiveLearning
+
+active_learner = ActiveLearning(ndim, bounds, data, run_simulator, prior, likelihood,
+ lhs_sampler, scalar_gasp)
+
+# %%md
+#
+# Then we can call the :py:meth:`.ActiveLearning.initial_simulation` method to run initial
+# simulations and call the :py:meth:`.ActiveLearning.iterative_emulation` method to
+# iteratively run new simulation and build emulator. Here we allocate 40 simulations for
+# initial emulator training and 60 simulations for adaptive training.
+#
+n0 = 40
+niter = 60
+
+init_var_samples, init_sim_outputs = active_learner.initial_simulation(
+ n0, mode='parallel', max_workers=4)
+
+var_samples, _, _ = active_learner.iterative_emulation(
+ n0, init_var_samples, init_sim_outputs, niter=niter)
+
+# %%md
+#
+# Once the active learning process is finished, we obtain the final emulator for the
+# logarithm of the unnormalized posterior, which is given by the
+# :py:meth:`.ActiveLearning.approx_ln_pxl` method.
+#
+# We can then estimate the posterior using grid estimation or Metropolis Hastings
+# estimation based on the emulator. An example is as follows. The contour plot shows
+# the estimated posterior.
+#
+from psimpy.inference import GridEstimation
+import matplotlib.pyplot as plt
+
+grid_estimator = GridEstimation(ndim, bounds, ln_pxl=active_learner.approx_ln_pxl)
+posterior, x_ndim = grid_estimator.run(nbins=50)
+
+fig, ax = plt.subplots(1,1,figsize=(5,4))
+
+# initial training points
+ax.scatter(init_var_samples[:,0], init_var_samples[:,1], s=10, c='r', marker='o',
+ zorder=1, alpha=0.8, label='initial training points')
+# actively picked training points
+ax.scatter(var_samples[n0:,0], var_samples[n0:,1], s=15, c='k', marker='+',
+ zorder=2, alpha=0.8, label='iterative training points')
+
+# estimated posterior based on the final emulator
+posterior = np.where(posterior < 1e-10, None, posterior)
+contour = ax.contour(x_ndim[0], x_ndim[1], np.transpose(posterior), levels=10, zorder=0)
+plt.colorbar(contour, ax=ax)
+
+ax.legend()
+ax.set_title('Active learning')
+ax.set_xlabel('x1')
+ax.set_ylabel('x2')
+ax.set_xlim([-5,5])
+ax.set_ylim([-5,5])
+plt.tight_layout()
\ No newline at end of file
--
GitLab