January 17, 2020

Probability Distributions for Schedule Risk Analysis

Here’s a post about some of the fundamental probability distributions used in Schedule Risk Analysis. After understanding these distributions and how to code them up in Python, the power of improving your project schedules is at your finger tips!

Check it out on github Last updated: 20/01/2022 22:58:37

Schedule Risk Analysis Distributions¶

The following are some distributions commonly used in Schedule Risk Analysis (SRA).

In this notebook, we will create classes to represet each distribution. In general, the classes will store the parameters, and three methods are available.

pdf to compute the probability of a sample, x
cdf to compute the cummulative distribution within a range a (low end) and b (high end)
rvs to sample from the distribution

Of particular interest is the rvs method, which may be used in Monte Carlo simulation in SRA. Metrics such as

Criticality Index (CI): measures the probability that an activity is on the critical path
Significance Index (SI): measures the relative importance of an activity
Schedule Sensitivity Index (SSI): measures the relative imporance of an activity taking CI into account
Cruciality Index (CRI): measures the correlation between the activity duration and total project duration

Triangular¶

The Triangular distribution is sometimes called a three-point estimate since it requires 3 parameters to create the distribution and the distribution's probability density shape ends up looking like a triangle. The parameters are

a minimum duration,
m expected duration, and
b maximum duration.

The probability density function (PDF) for the triangular distribution is as follows.

$f(x) = \begin{cases} 0 & \text{for } x < a, \\ \frac{2(x-a)}{(b-a)(c-a)} & \text{for } a \le x < c, \\ \frac{2}{b-a} & \text{for } x = c, \\ \frac{2(b-x)}{(b-a)(b-c)} & \text{for } c < x \le b, \\ 0 & \text{for } b < x. \end{cases}$

In [1]:

%matplotlib inline
import numpy as np
import random
import matplotlib.pyplot as plt
from numpy.random import uniform

random.seed(37)
np.random.seed(37)
plt.style.use('ggplot')

class Triangle(object):
    def __init__(self, a, m, b):
        self.a = a
        self.m = m
        self.b = b
    
    def pdf(self, x):
        a, m, b = self.a, self.m, self.b
        
        if b < x < a:
            return 0.0
        elif a <= x < m:
            return (2 * (x - a)) / ((b - a) * (m - a))
        elif x == m:
            return (2 / (b - a))
        elif m < x <= b:
            return (2 * (b - x)) / ((b - a) * (b - m))
        raise Exception(f'No conditions satisifed x={x}, {a}, {m}, {b}')
        
    def cdf(self):
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c
    
    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)
    
a, m, b = 500, 550, 600

d = Triangle(a, m, b)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

Beta¶

The Beta distribution is smoother and curved compared to the Triangular distribution. It is parameterized by the following.

a minimum duration
b maximum duration
$\theta_1$ shape parameter
$\theta_2$ shape parameter

Note that $\theta_1$ and $\theta_2$ are positive and modified to make the the distribution left ($\theta_1 > \theta_2$) or right skewed ($\theta_1 < \theta_2$). The PDF of the Beta distribution is defined as follows.

$f(x) = \phi \frac{\Gamma(\theta_1 + \theta_2)}{\Gamma(\theta_1) \Gamma(\theta_2) (b - a)^{\theta_1 + \theta_2 - 1}}$

In [2]:

from scipy.special import gamma

class Beta(object):
    def __init__(self, a, b, theta_1, theta_2):
        self.a = a
        self.b = b
        self.theta_1 = theta_1
        self.theta_2 = theta_2
        
    def pdf(self, x):
        a, b, theta_1, theta_2 = self.a, self.b, self.theta_1, self.theta_2
        
        p = gamma(theta_1 + theta_2)
        p = p / gamma(theta_1)
        p = p / gamma(theta_2)
        p = p / (b - a)**(theta_1 + theta_2 - 1)
        p = p * (x - a)**(theta_1 - 1)
        p = p * (b - x)**(theta_2 - 1)
        return p

    def cdf(self):
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c

    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)
    
a, b = 500, 600
theta_1, theta_2 = 6, 3

d = Beta(a, b, theta_1, theta_2)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

Beta Rectangular¶

The Beta Rectangular distribution is a mixture distribution of the Beta and Uniform distributions. It allows for an estimator's judgement to vary between Beta and complete uncertainty (Uniform). The parameters for a Beta Rectangular distribution are as follows.

a minimum duration
b maximum duration
$\theta_1$ shape parameter
$\theta_2$ shape parameter
$\theta$ mixing parameter

Note that when $\theta=1$, then the Beta distribution is recovered, and when $\theta=0$, then the Uniform distribution is recovered. The parameter $\theta$ is called the mixing parameter. The PDF of the Beta Rectangular is defined as follows.

$f(x) = \phi \frac{\Gamma(\theta_1 + \theta_2)}{\Gamma(\theta_1) \Gamma(\theta_2) (b - a)^{\theta_1 + \theta_2 - 1}} + \frac{1 - \theta}{b - a}$

In [3]:

class BetaRect(object):
    def __init__(self, a, b, theta_1, theta_2, theta):
        self.a = a
        self.b = b
        self.theta_1 = theta_1
        self.theta_2 = theta_2
        self.theta = theta
        
    def pdf(self, x):
        a, b = self.a, self.b
        theta_1, theta_2, theta = self.theta_1, self.theta_2, self.theta
        
        p = gamma(theta_1 + theta_2)
        p = p / gamma(theta_1)
        p = p / gamma(theta_2)
        p = p / (b - a)**(theta_1 + theta_2 - 1)
        p = p * (x - a)**(theta_1 - 1)
        p = p * (b - x)**(theta_2 - 1)
        p = p * theta
        p = p + ((1 - theta) / (b - a))
        return p

    def cdf(self):
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c

    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)

a, b = 500, 600
theta_1, theta_2 = 6, 3
theta = 0.5

d = BetaRect(a, b, theta_1, theta_2, theta)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

Doubly-truncated¶

Compared to the Beta distribution, the Doubly-truncated distribution emphasizes more on the mode. It is parameterized by the following.

a minimum duration
b maximum duration
$\mu$ mean
$\sigma$ standard deviation

The PDF of the Doubly-truncated distribution is defined as follows.

$f(x)=\frac{\phi(\frac{x - \mu}{\sigma})}{\sigma( \Phi(\frac{b - \mu}{\sigma}) - \Phi(\frac{b - \mu}{\sigma}) )}$

In [4]:

from scipy.stats import norm

class DoublyTruncated(object):
    def __init__(self, a, b, mu, sigma):
        self.a = a
        self.b = b
        self.mu = mu
        self.sigma = sigma
        
    def pdf(self, x):
        a, b = self.a, self.b
        mu, sigma = self.mu, self.sigma
        
        n = norm.pdf((x - mu) / sigma)
        d1 = norm.cdf((b - mu) / sigma)
        d2 = norm.cdf((a - mu) / sigma)
        d = sigma * (d1 - d2)
        p = n / d
        return p

    def cdf(self):
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c

    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)

a, b = 500, 600
mu, sigma = 570, 10

d = DoublyTruncated(a, b, mu, sigma)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

Log-normal¶

The Log-normal distribution is used in SRA when activities are right-skewed. This distribution has the following parameters.

$\mu$ mean
$\sigma$ standard deviation

The PDF of the Log-normal distribution is defined as follows.

$f(x)=\frac{1}{x} \frac{1}{\sigma\sqrt{2\pi}} \exp\left( -\frac{(\ln x-\mu)^2}{2\sigma^2}\right)$

In [5]:

from scipy.stats import lognorm

class LogNormal(object):
    def __init__(self, a, b, s, mu, sigma):
        self.a = a
        self.b = b
        self.s = s
        self.mu = mu
        self.sigma = sigma
        
    def pdf(self, x):
        s, mu, sigma = self.s, self.mu, self.sigma
        
        return lognorm.pdf(x, s, loc=mu, scale=sigma)

    def cdf(self):
        a, b = self.a, self.b
        
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c

    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)

a, b = 500, 600
mu, sigma, s = 500, 10, 0.9

d = LogNormal(a, b, s, mu, sigma)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

Weibull¶

Compared to the Log-normal distribution, the Weibull distribution is used when activity durations are left-skewed. Its parameters are as follows.

$\mu$ mean
$\sigma$ standard deviation

The PDF of the Weibull is defined as follows.

$f(x;\lambda,k) = \begin{cases} \frac{k}{\lambda}\left(\frac{x}{\lambda}\right)^{k-1}e^{-(x/\lambda)^{k}} & x\geq0 ,\\ 0 & x<0, \end{cases}$

In [6]:

from scipy.stats import weibull_min

class Weibull(object):
    def __init__(self, a, b, s, mu, sigma):
        self.a = a
        self.b = b
        self.s = s
        self.mu = mu
        self.sigma = sigma
        
    def pdf(self, x):
        s, mu, sigma = self.s, self.mu, self.sigma
        
        return weibull_min.pdf(x, s, loc=mu, scale=sigma)

    def cdf(self):
        a, b = self.a, self.b
        
        p = np.array([self.pdf(x) for x in range(a, b+1)])
        c = p.cumsum()
        return c

    def rvs(self, size=1):
        a, b = self.a, self.b
        
        o = {i:v for i, v in enumerate(range(a, b+1))}
        c = self.cdf()
        u = uniform(size=size)
        s = [o[np.argmax(c >= x)] for x in u]
        return np.array(s)

a, b = 500, 600
mu, sigma, s = 500, 10, 5

d = Weibull(a, b, s, mu, sigma)

p = [d.pdf(x) for x in range(a, b+1)]
c = d.cdf()
s = d.rvs(size=20000)

fig, ax = plt.subplots(1, 3, figsize=(15, 4))
_ = ax[0].plot(list(range(a, b+1)), p)
_ = ax[1].plot(list(range(a, b+1)), c)
_ = ax[2].hist(s, bins=list(range(a, b+1)), density=True)

oneoffcoder