Non parametric Adaptive Importance Sampling (NAIS)

The objective is to evaluate a probability from the Non parametric Adaptive Importance Sampling (NAIS) technique.

We consider the four-branch function g : \mathbb{R}^2 \rightarrow \mathbb{R} defined by:

\begin{align*}
g(\vect{X}) = \min \begin{pmatrix}5+0.1(x_1-x_2)^2-\frac{(x_1+x_2)}{\sqrt{2}}\\
5+0.1(x_1-x_2)^2+\frac{(x_1+x_2)}{\sqrt{2}}\\
(x_1-x_2)+ \frac{9}{\sqrt{2}}\\
(x_2-x_1)+ \frac{9}{\sqrt{2}}
\end{pmatrix}
\end{align*}

and the input random vector \vect{X} = (X_1, X_2) which follows the standard 2-dimensional Normal distribution:

\begin{align*}
\vect{X} \sim  \mathcal{N}(\mu = [0, 0], \sigma = [1,1], corr = \mat{I}_2)
\end{align*}

We want to evaluate the probability:

\begin{align*}
p = \mathbb{P} ( g(\vect{X}) \leq 0 )
\end{align*}

First, import the python modules:

import openturns as ot
from openturns.viewer import View
import math

Create the probabilistic model Y = g(\vect{X})

Create the input random vector \vect{X}:

X = ot.RandomVector(ot.Normal(2))

Create the function g from a PythonFunction:

def fourBranch(x):
    x1 = x[0]
    x2 = x[1]

    g1 = 5 + 0.1 * (x1 - x2) ** 2 - (x1 + x2) / math.sqrt(2)
    g2 = 5 + 0.1 * (x1 - x2) ** 2 + (x1 + x2) / math.sqrt(2)
    g3 = (x1 - x2) + 9 / math.sqrt(2)
    g4 = (x2 - x1) + 9 / math.sqrt(2)

    return [min((g1, g2, g3, g4))]


g = ot.PythonFunction(2, 1, fourBranch)

Draw the function g to help to understand the shape of the limit state function:

graph = ot.Graph("Four Branch function", "x1", "x2", True, "upper right")
drawfunction = g.draw([-8] * 2, [8] * 2, [100] * 2)
graph.add(drawfunction)
view = View(graph)
Four Branch function

In order to be able to get the NAIS samples used in the algorithm, it is necessary to transform the PythonFunction into a MemoizeFunction:

g = ot.MemoizeFunction(g)

Create the output random vector Y = g(\vect{X}):

Y = ot.CompositeRandomVector(g, X)

Create the event \{ Y = g(\vect{X}) \leq 0 \}

threshold = 0.0
myEvent = ot.ThresholdEvent(Y, ot.Less(), threshold)

Evaluate the probability with the NAIS technique

quantileLevel = 0.1
algo = ot.NAIS(myEvent, quantileLevel)

Now you can run the algorithm.

algo.run()
result = algo.getResult()
proba = result.getProbabilityEstimate()
print("Proba NAIS = ", proba)
print("Current coefficient of variation = ", result.getCoefficientOfVariation())
Proba NAIS =  8.001913221123705e-06
Current coefficient of variation =  0.09906223602270714

The length of the confidence interval of level 95\% is:

length95 = result.getConfidenceLength()
print("Confidence length (0.95) = ", result.getConfidenceLength())
Confidence length (0.95) =  3.1072775732814027e-06

which enables to build the confidence interval:

print(
    "Confidence interval (0.95) = [",
    proba - length95 / 2,
    ", ",
    proba + length95 / 2,
    "]",
)
Confidence interval (0.95) = [ 6.4482744344830036e-06 ,  9.555552007764406e-06 ]

Draw the NAIS samples used by the algorithm

The following manipulations are possible only if you have created a MemoizeFunction that enables to store all the inputs and outputs of the function g.

Get all the inputs and outputs where g were evaluated:

inputNAIS = g.getInputHistory()
outputNAIS = g.getOutputHistory()
nTotal = inputNAIS.getSize()
print("Number of evaluations of g = ", nTotal)
Number of evaluations of g =  4000

Within each step of the algorithm, a sample of size N is created, where:

N = algo.getMaximumOuterSampling() * algo.getBlockSize()
print("Size of each subset = ", N)
Size of each subset =  1000

You can get the number N_s of steps with:

Ns = int(nTotal / N)
print("Number of steps = ", Ns)
Number of steps =  4

Now, we can split the initial sample into NAIS samples of size N_s:

listNAISSamples = list()
listOutputNAISSamples = list()
for i in range(Ns):
    listNAISSamples.append(inputNAIS[i * N: i * N + N])
    listOutputNAISSamples.append(outputNAIS[i * N: i * N + N])

And get all the levels defining the intermediate and final thresholds given by the empirical quantiles of each NAIS output sample:

levels = []
for i in range(Ns - 1):
    levels.append(listOutputNAISSamples[i].computeQuantile(quantileLevel)[0])
levels.append(threshold)

The following graph draws each NAIS sample and the frontier g(x_1, x_2) = l_i where l_i is the threshold at the step i:

graph = ot.Graph("NAIS samples", "x1", "x2", True, "lower left")
graph.setGrid(True)

Add all the NAIS samples:

for i in range(Ns):
    cloud = ot.Cloud(listNAISSamples[i])
    graph.add(cloud)
col = ot.Drawable().BuildDefaultPalette(Ns)
graph.setColors(col)

Add the frontiers g(x_1, x_2) = l_i where l_i is the threshold at the step i:

gIsoLines = g.draw([-8] * 2, [8] * 2, [128] * 2)
dr = gIsoLines.getDrawable(2)
for i, lv in enumerate(levels):
    dr.setLevels([lv])
    dr.setLineStyle("solid")
    dr.setLegend(r"$g(X) = $" + str(round(lv, 2)))
    dr.setLineWidth(3)
    dr.setColor(col[i])
    graph.add(dr)

# sphinx_gallery_thumbnail_number = 2
view = View(graph)
NAIS samples

Draw the frontiers only

The following graph enables to understand the progression of the algorithm from the mean value of the initial distribution to the limit state function:

graph = ot.Graph("NAIS thresholds", "x1", "x2", True, "lower left")
graph.setGrid(True)
dr = gIsoLines.getDrawable(0)
for i, lv in enumerate(levels):
    dr.setLevels([lv])
    dr.setLineStyle("solid")
    dr.setLegend(r"$g(X) = $" + str(round(lv, 2)))
    dr.setLineWidth(3)
    graph.add(dr)

graph.setColors(col)
view = View(graph)
NAIS thresholds