Skip to content

A Python implementation of the Analytic Hierarchy Process

License

Notifications You must be signed in to change notification settings

PhilipGriffith/AHPy

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AHPy

AHPy is an implementation of the Analytic Hierarchy Process (AHP), a method used to structure, synthesize and evaluate the elements of a decision problem. Developed by Thomas Saaty in the 1970s, AHP's broad use in fields well beyond that of operational research is a testament to its simple yet powerful combination of psychology and mathematics.

AHPy attempts to provide a library that is not only simple to use, but also capable of intuitively working within the numerous conceptual frameworks to which the AHP can be applied. For this reason, general terms have been preferred to more specific ones within the programming interface.

Installing AHPy

AHPy is available on the Python Package Index (PyPI):

python -m pip install ahpy

AHPy requires Python 3.7+, as well as numpy and scipy.

Table of Contents

Examples

Relative consumption of drinks in the United States

Choosing a leader

Purchasing a vehicle

Purchasing a vehicle reprised: normalized weights and the Compose class

Details

The Compare Class

Compare.add_children()

Compare.report()

The Compose Class

Compose.add_comparisons()

Compose.add_hierarchy()

Compose.report()

A Note on Weights

Missing Pairwise Comparisons

Development and Testing


Examples

The easiest way to learn how to use AHPy is to see it used, so this README begins with worked examples of gradually increasing complexity.

Relative consumption of drinks in the United States

This example is often used in Saaty's expositions of the AHP as a brief but clear demonstration of the method; it's what first opened my eyes to the broad usefulness of the AHP (as well as the wisdom of crowds!). The version I'm using here is from his 2008 article 'Decision making with the analytic hierarchy process'. If you're unfamiliar with the example, 30 participants were asked to compare the relative consumption of drinks in the United States. For instance, they believed that coffee was consumed much more than wine, but at the same rate as milk. The matrix derived from their answers was as follows:

Coffee Wine Tea Beer Soda Milk Water
Coffee 1 9 5 2 1 1 1/2
Wine 1/9 1 1/3 1/9 1/9 1/9 1/9
Tea 1/5 3 1 1/3 1/4 1/3 1/9
Beer 1/2 9 3 1 1/2 1 1/3
Soda 1 9 4 2 1 2 1/2
Milk 1 9 3 1 1/2 1 1/3
Water 2 9 9 3 2 3 1

The table below shows the relative consumption of drinks as computed using the AHP, given this matrix, together with the actual relative consumption of drinks as obtained from U.S. Statistical Abstracts:

🤯 Coffee Wine Tea Beer Soda Milk Water
AHP 0.177 0.019 0.042 0.116 0.190 0.129 0.327
Actual 0.180 0.010 0.040 0.120 0.180 0.140 0.330

We can recreate this analysis with AHPy using the following code:

>>> drink_comparisons = {('coffee', 'wine'): 9, ('coffee', 'tea'): 5, ('coffee', 'beer'): 2, ('coffee', 'soda'): 1,
			 ('coffee', 'milk'): 1, ('coffee', 'water'): 1 / 2,
                         ('wine', 'tea'): 1 / 3, ('wine', 'beer'): 1 / 9, ('wine', 'soda'): 1 / 9,
                         ('wine', 'milk'): 1 / 9, ('wine', 'water'): 1 / 9,
                         ('tea', 'beer'): 1 / 3, ('tea', 'soda'): 1 / 4, ('tea', 'milk'): 1 / 3,
                         ('tea', 'water'): 1 / 9,
                         ('beer', 'soda'): 1 / 2, ('beer', 'milk'): 1, ('beer', 'water'): 1 / 3,
                         ('soda', 'milk'): 2, ('soda', 'water'): 1 / 2,
                         ('milk', 'water'): 1 / 3}

>>> drinks = ahpy.Compare(name='Drinks', comparisons=drink_comparisons, precision=3, random_index='saaty')

>>> print(drinks.target_weights)
{'water': 0.327, 'soda': 0.19, 'coffee': 0.177, 'milk': 0.129, 'beer': 0.116, 'tea': 0.042, 'wine': 0.019}

>>> print(drinks.consistency_ratio)
0.022
  1. First, we create a dictionary of pairwise comparisons using the values from the matrix above.
  2. We then create a Compare object, initializing it with a unique name and the dictionary we just made. We also change the precision and random index so that the results match those provided by Saaty.
  3. Finally, we print the Compare object's target weights and consistency ratio to see the results of our analysis.

Brilliant!

Choosing a leader

This example can be found in an appendix to the Wikipedia entry for AHP. The names have been changed in a nod to the original saying, but the input comparison values remain the same.

N.B.

You may notice that in some cases AHPy's results will not match those on the Wikipedia page. This is not an error in AHPy's calculations, but rather a result of the method used to compute the values shown in the Wikipedia examples:

You can duplicate this analysis at this online demonstration site...IMPORTANT: The demo site is designed for convenience, not accuracy. The priorities it returns may differ somewhat from those returned by rigorous AHP calculations.

In this example, we'll be judging job candidates by their experience, education, charisma and age. Therefore, we need to compare each potential leader to the others, given each criterion...

>>> experience_comparisons = {('Moll', 'Nell'): 1/4, ('Moll', 'Sue'): 4, ('Nell', 'Sue'): 9}
>>> education_comparisons = {('Moll', 'Nell'): 3, ('Moll', 'Sue'): 1/5, ('Nell', 'Sue'): 1/7}
>>> charisma_comparisons = {('Moll', 'Nell'): 5, ('Moll', 'Sue'): 9, ('Nell', 'Sue'): 4}
>>> age_comparisons = {('Moll', 'Nell'): 1/3, ('Moll', 'Sue'): 5, ('Nell', 'Sue'): 9}

...as well as compare the importance of each criterion to the others:

>>> criteria_comparisons = {('Experience', 'Education'): 4, ('Experience', 'Charisma'): 3, ('Experience', 'Age'): 7,
			    ('Education', 'Charisma'): 1/3, ('Education', 'Age'): 3,
			    ('Charisma', 'Age'): 5}

Before moving on, it's important to note that the order of the elements that form the dictionaries' keys is meaningful. For example, using Saaty's scale, the comparison ('Experience', 'Education'): 4 means that "Experience is moderately+ more important than Education."

Now that we've created all of the necessary pairwise comparison dictionaries, we'll create their corresponding Compare objects and use the dictionaries as input:

>>> experience = ahpy.Compare('Experience', experience_comparisons, precision=3, random_index='saaty')
>>> education = ahpy.Compare('Education', education_comparisons, precision=3, random_index='saaty')
>>> charisma = ahpy.Compare('Charisma', charisma_comparisons, precision=3, random_index='saaty')
>>> age = ahpy.Compare('Age', age_comparisons, precision=3, random_index='saaty')
>>> criteria = ahpy.Compare('Criteria', criteria_comparisons, precision=3, random_index='saaty')

Notice that the names of the Experience, Education, Charisma and Age objects are repeated in the criteria_comparisons dictionary above. This is necessary in order to properly link the Compare objects together into a hierarchy, as shown next.

In the final step, we need to link the Compare objects together into a hierarchy, such that Criteria is the parent object and the other objects form its children:

>>> criteria.add_children([experience, education, charisma, age])

Now that the hierarchy represents the decision problem, we can print the target weights of the parent Criteria object to see the results of the analysis:

>>> print(criteria.target_weights)
{'Nell': 0.493, 'Moll': 0.358, 'Sue': 0.15}

We can also print the local and global weights of the elements within any of the other Compare objects, as well as the consistency ratio of their comparisons:

>>> print(experience.local_weights)
{'Nell': 0.717, 'Moll': 0.217, 'Sue': 0.066}

>>> print(experience.consistency_ratio)
0.035

>>> print(education.global_weights)
{'Sue': 0.093, 'Moll': 0.024, 'Nell': 0.01}

>>> print(education.consistency_ratio)
0.062

The global and local weights of the Compare objects themselves are likewise available:

>>> print(experience.global_weight)
0.548

>>> print(education.local_weight)
0.127

Calling report() on a Compare object provides a standard way to learn information about the object. In the code below, the variable report contains a Python dictionary of important information, while the show=True argument prints the same information to the console in JSON format:

>>> report = criteria.report(show=True)
{
    "Criteria": {
        "global_weight": 1.0,
        "local_weight": 1.0,
        "target_weights": {
            "Nell": 0.493,
            "Moll": 0.358,
            "Sue": 0.15
        },
        "elements": {
            "global_weights": {
                "Experience": 0.548,
                "Charisma": 0.27,
                "Education": 0.127,
                "Age": 0.056
            },
            "local_weights": {
                "Experience": 0.548,
                "Charisma": 0.27,
                "Education": 0.127,
                "Age": 0.056
            },
            "consistency_ratio": 0.044
        }
    }
}

Purchasing a vehicle

This example can also be found in an appendix to the Wikipedia entry for AHP. Like before, in some cases AHPy's results will not match those on the Wikipedia page, even though the input comparison values are identical. To reiterate, this is due to a difference in methods, not an error in AHPy.

In this example, we'll be choosing a vehicle to purchase based on its cost, safety, style and capacity. Cost will further depend on a combination of the vehicle's purchase price, fuel costs, maintenance costs and resale value; capacity will depend on a combination of the vehicle's cargo and passenger capacity.

First, we compare the high-level criteria to one another:

>>> criteria_comparisons = {('Cost', 'Safety'): 3, ('Cost', 'Style'): 7, ('Cost', 'Capacity'): 3,
			    ('Safety', 'Style'): 9, ('Safety', 'Capacity'): 1,
			    ('Style', 'Capacity'): 1/7}

If we create a Compare object for the criteria, we can view its report:

>>> criteria = ahpy.Compare('Criteria', criteria_comparisons, precision=3)
>>> report = criteria.report(show=True)
{
    "Criteria": {
        "global_weight": 1.0,
        "local_weight": 1.0,
        "target_weights": {
            "Cost": 0.51,
            "Safety": 0.234,
            "Capacity": 0.215,
            "Style": 0.041
        },
        "elements": {
            "global_weights": {
                "Cost": 0.51,
                "Safety": 0.234,
                "Capacity": 0.215,
                "Style": 0.041
            },
            "local_weights": {
                "Cost": 0.51,
                "Safety": 0.234,
                "Capacity": 0.215,
                "Style": 0.041
            },
            "consistency_ratio": 0.08
        }
    }
}

Next, we compare the subcriteria of Cost to one another...

>>> cost_comparisons = {('Price', 'Fuel'): 2, ('Price', 'Maintenance'): 5, ('Price', 'Resale'): 3,
			('Fuel', 'Maintenance'): 2, ('Fuel', 'Resale'): 2,
			('Maintenance', 'Resale'): 1/2}

...as well as the subcriteria of Capacity:

>>> capacity_comparisons = {('Cargo', 'Passenger'): 1/5}

We also need to compare each of the potential vehicles to the others, given each criterion. We'll begin by building a list of all possible two-vehicle combinations:

>>> import itertools
>>> vehicles = ('Accord Sedan', 'Accord Hybrid', 'Pilot', 'CR-V', 'Element', 'Odyssey')
>>> vehicle_pairs = list(itertools.combinations(vehicles, 2))
>>> print(vehicle_pairs)
[('Accord Sedan', 'Accord Hybrid'), ('Accord Sedan', 'Pilot'), ('Accord Sedan', 'CR-V'), ('Accord Sedan', 'Element'), ('Accord Sedan', 'Odyssey'), ('Accord Hybrid', 'Pilot'), ('Accord Hybrid', 'CR-V'), ('Accord Hybrid', 'Element'), ('Accord Hybrid', 'Odyssey'), ('Pilot', 'CR-V'), ('Pilot', 'Element'), ('Pilot', 'Odyssey'), ('CR-V', 'Element'), ('CR-V', 'Odyssey'), ('Element', 'Odyssey')]

Then we can simply zip together the vehicle pairs and their pairwise comparison values for each criterion:

>>> price_values = (9, 9, 1, 1/2, 5, 1, 1/9, 1/9, 1/7, 1/9, 1/9, 1/7, 1/2, 5, 6)
>>> price_comparisons = dict(zip(vehicle_pairs, price_values))
>>> print(price_comparisons)
{('Accord Sedan', 'Accord Hybrid'): 9, ('Accord Sedan', 'Pilot'): 9, ('Accord Sedan', 'CR-V'): 1, ('Accord Sedan', 'Element'): 0.5, ('Accord Sedan', 'Odyssey'): 5, ('Accord Hybrid', 'Pilot'): 1, ('Accord Hybrid', 'CR-V'): 0.1111111111111111, ('Accord Hybrid', 'Element'): 0.1111111111111111, ('Accord Hybrid', 'Odyssey'): 0.14285714285714285, ('Pilot', 'CR-V'): 0.1111111111111111, ('Pilot', 'Element'): 0.1111111111111111, ('Pilot', 'Odyssey'): 0.14285714285714285, ('CR-V', 'Element'): 0.5, ('CR-V', 'Odyssey'): 5, ('Element', 'Odyssey'): 6}

>>> safety_values = (1, 5, 7, 9, 1/3, 5, 7, 9, 1/3, 2, 9, 1/8, 2, 1/8, 1/9)
>>> safety_comparisons = dict(zip(vehicle_pairs, safety_values))

>>> passenger_values = (1, 1/2, 1, 3, 1/2, 1/2, 1, 3, 1/2, 2, 6, 1, 3, 1/2, 1/6)
>>> passenger_comparisons = dict(zip(vehicle_pairs, passenger_values))

>>> fuel_values = (1/1.13, 1.41, 1.15, 1.24, 1.19, 1.59, 1.3, 1.4, 1.35, 1/1.23, 1/1.14, 1/1.18, 1.08, 1.04, 1/1.04)
>>> fuel_comparisons = dict(zip(vehicle_pairs, fuel_values))

>>> resale_values = (3, 4, 1/2, 2, 2, 2, 1/5, 1, 1, 1/6, 1/2, 1/2, 4, 4, 1)
>>> resale_comparisons = dict(zip(vehicle_pairs, resale_values))

>>> maintenance_values = (1.5, 4, 4, 4, 5, 4, 4, 4, 5, 1, 1.2, 1, 1, 3, 2)
>>> maintenance_comparisons = dict(zip(vehicle_pairs, maintenance_values))

>>> style_values = (1, 7, 5, 9, 6, 7, 5, 9, 6, 1/6, 3, 1/3, 7, 5, 1/5)
>>> style_comparisons = dict(zip(vehicle_pairs, style_values))

>>> cargo_values = (1, 1/2, 1/2, 1/2, 1/3, 1/2, 1/2, 1/2, 1/3, 1, 1, 1/2, 1, 1/2, 1/2)
>>> cargo_comparisons = dict(zip(vehicle_pairs, cargo_values))

Now that we've created all of the necessary pairwise comparison dictionaries, we can create their corresponding Compare objects:

>>> cost = ahpy.Compare('Cost', cost_comparisons, precision=3)
>>> capacity = ahpy.Compare('Capacity', capacity_comparisons, precision=3)
>>> price = ahpy.Compare('Price', price_comparisons, precision=3)
>>> safety = ahpy.Compare('Safety', safety_comparisons, precision=3)
>>> passenger = ahpy.Compare('Passenger', passenger_comparisons, precision=3)
>>> fuel = ahpy.Compare('Fuel', fuel_comparisons, precision=3)
>>> resale = ahpy.Compare('Resale', resale_comparisons, precision=3)
>>> maintenance = ahpy.Compare('Maintenance', maintenance_comparisons, precision=3)
>>> style = ahpy.Compare('Style', style_comparisons, precision=3)
>>> cargo = ahpy.Compare('Cargo', cargo_comparisons, precision=3)

The final step is to link all of the Compare objects into a hierarchy. First, we'll make the Price, Fuel, Maintenance and Resale objects the children of the Cost object...

>>> cost.add_children([price, fuel, maintenance, resale])

...and do the same to link the Cargo and Passenger objects to the Capacity object...

>>> capacity.add_children([cargo, passenger])

...then finally make the Cost, Safety, Style and Capacity objects the children of the Criteria object:

>>> criteria.add_children([cost, safety, style, capacity])

Now that the hierarchy represents the decision problem, we can print the target weights of the highest level Criteria object to see the results of the analysis:

>>> print(criteria.target_weights)
{'Odyssey': 0.219, 'Accord Sedan': 0.215, 'CR-V': 0.167, 'Accord Hybrid': 0.15, 'Element': 0.144, 'Pilot': 0.106}

For detailed information about any of the Compare objects in the hierarchy, we can call that object's report() with the verbose=True argument:

>>> report = criteria.report(show=True, verbose=True)
{
    "name": "Criteria",
    "global_weight": 1.0,
    "local_weight": 1.0,
    "target_weights": {
        "Odyssey": 0.219,
        "Accord Sedan": 0.215,
        "CR-V": 0.167,
        "Accord Hybrid": 0.15,
        "Element": 0.144,
        "Pilot": 0.106
    },
    "elements": {
        "global_weights": {
            "Cost": 0.51,
            "Safety": 0.234,
            "Capacity": 0.215,
            "Style": 0.041
        },
        "local_weights": {
            "Cost": 0.51,
            "Safety": 0.234,
            "Capacity": 0.215,
            "Style": 0.041
        },
        "consistency_ratio": 0.08,
        "random_index": "Donegan & Dodd",
        "count": 4,
        "names": [
            "Cost",
            "Safety",
            "Style",
            "Capacity"
        ]
    },
    "children": {
        "count": 4,
        "names": [
            "Cost",
            "Safety",
            "Style",
            "Capacity"
        ]
    },
    "comparisons": {
        "count": 6,
        "input": {
            "Cost, Safety": 3,
            "Cost, Style": 7,
            "Cost, Capacity": 3,
            "Safety, Style": 9,
            "Safety, Capacity": 1,
            "Style, Capacity": 0.14285714285714285
        },
        "computed": null
    }
}

Calling report(show=True, verbose=True) on Compare objects at lower levels of the hierarchy will provide different information, depending on the level they're in:

>>> report = cost.report(show=True, verbose=True)
{
    "name": "Cost",
    "global_weight": 0.51,
    "local_weight": 0.51,
    "target_weights": null,
    "elements": {
        "global_weights": {
            "Price": 0.249,
            "Fuel": 0.129,
            "Resale": 0.082,
            "Maintenance": 0.051
        },
        "local_weights": {
            "Price": 0.488,
            "Fuel": 0.252,
            "Resale": 0.161,
            "Maintenance": 0.1
        },
        "consistency_ratio": 0.016,
        "random_index": "Donegan & Dodd",
        "count": 4,
        "names": [
            "Price",
            "Fuel",
            "Maintenance",
            "Resale"
        ]
    },
    "children": {
        "count": 4,
        "names": [
            "Price",
            "Fuel",
            "Resale",
            "Maintenance"
        ]
    },
    "comparisons": {
        "count": 6,
        "input": {
            "Price, Fuel": 2,
            "Price, Maintenance": 5,
            "Price, Resale": 3,
            "Fuel, Maintenance": 2,
            "Fuel, Resale": 2,
            "Maintenance, Resale": 0.5
        },
        "computed": null
    }
}

>>> report = price.report(show=True, verbose=True)
{
    "name": "Price",
    "global_weight": 0.249,
    "local_weight": 0.488,
    "target_weights": null,
    "elements": {
        "global_weights": {
            "Element": 0.091,
            "Accord Sedan": 0.061,
            "CR-V": 0.061,
            "Odyssey": 0.023,
            "Accord Hybrid": 0.006,
            "Pilot": 0.006
        },
        "local_weights": {
            "Element": 0.366,
            "Accord Sedan": 0.246,
            "CR-V": 0.246,
            "Odyssey": 0.093,
            "Accord Hybrid": 0.025,
            "Pilot": 0.025
        },
        "consistency_ratio": 0.072,
        "random_index": "Donegan & Dodd",
        "count": 6,
        "names": [
            "Accord Sedan",
            "Accord Hybrid",
            "Pilot",
            "CR-V",
            "Element",
            "Odyssey"
        ]
    },
    "children": null,
    "comparisons": {
        "count": 15,
        "input": {
            "Accord Sedan, Accord Hybrid": 9,
            "Accord Sedan, Pilot": 9,
            "Accord Sedan, CR-V": 1,
            "Accord Sedan, Element": 0.5,
            "Accord Sedan, Odyssey": 5,
            "Accord Hybrid, Pilot": 1,
            "Accord Hybrid, CR-V": 0.1111111111111111,
            "Accord Hybrid, Element": 0.1111111111111111,
            "Accord Hybrid, Odyssey": 0.14285714285714285,
            "Pilot, CR-V": 0.1111111111111111,
            "Pilot, Element": 0.1111111111111111,
            "Pilot, Odyssey": 0.14285714285714285,
            "CR-V, Element": 0.5,
            "CR-V, Odyssey": 5,
            "Element, Odyssey": 6
        },
        "computed": null
    }
}

Finally, calling report(complete=True) on any Compare object in the hierarchy will return a dictionary containing a report for every Compare object in the hierarchy, with the keys of the dictionary being the names of the Compare objects:

>>> complete_report = cargo.report(complete=True)

>>> print([key for key in complete_report])
['Criteria', 'Cost', 'Price', 'Fuel', 'Maintenance', 'Resale', 'Safety', 'Style', 'Capacity', 'Cargo', 'Passenger']

>>> print(complete_report['Cargo'])
{'name': 'Cargo', 'global_weight': 0.0358, 'local_weight': 0.1667, 'target_weights': None, 'elements': {'global_weights': {'Odyssey': 0.011, 'Pilot': 0.006, 'CR-V': 0.006, 'Element': 0.006, 'Accord Sedan': 0.003, 'Accord Hybrid': 0.003}, 'local_weights': {'Odyssey': 0.311, 'Pilot': 0.17, 'CR-V': 0.17, 'Element': 0.17, 'Accord Sedan': 0.089, 'Accord Hybrid': 0.089}, 'consistency_ratio': 0.002}}

>>> print(complete_report['Criteria']['target_weights'])
{'Odyssey': 0.219, 'Accord Sedan': 0.215, 'CR-V': 0.167, 'Accord Hybrid': 0.15, 'Element': 0.144, 'Pilot': 0.106}

Calling report(complete=True, verbose=True) will return a similar dictionary, but with the detailed version of the reports.

>>> complete_report = style.report(complete=True, verbose=True)

>>> print(complete_report['Price']['comparisons']['count'])
15

We could also print all of the reports to the console with the show=True argument.

Purchasing a vehicle reprised: normalized weights and the Compose class

After reading through the explanation of the vehicle decision problem on Wikipedia, you may have wondered whether the data used to represent purely numeric criteria (such as passenger capacity) could be used directly when comparing the vehicles to one another, rather than requiring tranformation into judgments of "intensity." In this example, we'll solve the same decision problem as before, except now we'll normalize the measured values for passenger capacity, fuel costs, resale value and cargo capacity in order to arrive at a different set of weights for these criteria.

We'll also use a Compose object to structure the decision problem. The Compose object allows us to work with an abstract representation of the problem hierarchy, rather than build it dynamically with code, which is valuable when we're not using AHPy in an interactive setting. To use the Compose object, we'll need to first add the comparison information, then the hierarchy, in that order. But more on that later.

Using the list of vehicles from the previous example, we'll first zip together the vehicles and their measured values, then create a Compare object for each of our normalized criteria:

>>> passenger_measured_values = (5, 5, 8, 5, 4, 8)
>>> passenger_data = dict(zip(vehicles, passenger_measured_values))
>>> print(passenger_data)
{'Accord Sedan': 5, 'Accord Hybrid': 5, 'Pilot': 8, 'CR-V': 5, 'Element': 4, 'Odyssey': 8}

>>> passenger_normalized = ahp.Compare('Passenger', passenger_data, precision=3)

>>> fuel_measured_values = (31, 35, 22, 27, 25, 26)
>>> fuel_data = dict(zip(vehicles, fuel_measured_values))
>>> fuel_normalized = ahp.Compare('Fuel', fuel_data, precision=3)

>>> resale_measured_values = (0.52, 0.46, 0.44, 0.55, 0.48, 0.48)
>>> resale_data = dict(zip(vehicles, resale_measured_values))
>>> resale_normalized = ahp.Compare('Resale', resale_data, precision=3)

>>> cargo_measured_values = (14, 14, 87.6, 72.9, 74.6, 147.4)
>>> cargo_data = dict(zip(vehicles, cargo_measured_values))
>>> cargo_normalized = ahp.Compare('Cargo', cargo_data, precision=3)

Let's print the normalized local weights of the new Passenger object to compare them to the local weights in the previous example:

>>> print(passenger_normalized.local_weights)
{'Pilot': 0.229, 'Odyssey': 0.229, 'Accord Sedan': 0.143, 'Accord Hybrid': 0.143, 'CR-V': 0.143, 'Element': 0.114}

>>> print(passenger.local_weights)
{'Accord Sedan': 0.493, 'Accord Hybrid': 0.197, 'Odyssey': 0.113, 'Element': 0.091, 'CR-V': 0.057, 'Pilot': 0.049}

When we use the measured values directly, we see that the rankings for the vehicles are different than they were before. Whether this will affect the synthesized rankings of the target variables remains to be seen, however.

We next create a Compose object and begin to add the comparison information:

>>> compose = ahpy.Compose()

>>> compose.add_comparisons([passenger_normalized, fuel_normalized, resale_normalized, cargo_normalized])

We can add comparison information to the Compose object in a few different ways. As shown above, we can provide a list of Compare objects; we can also provide them one at a time or stored in a tuple. Using Compare objects from our previous example:

>>> compose.add_comparisons(cost)

>>> compose.add_comparisons((safety, style, capacity))

We can even treat the Compose object like a Compare object and add the data directly. Again using code from the previous example:

>>> compose.add_comparisons('Price', price_comparisons, precision=3)

Finally, we can provide an ordered list or tuple containing the data needed to construct a Compare object:

>>> comparisons = [('Maintenance', maintenance_comparisons, 3), ('Criteria', criteria_comparisons)]
>>> compose.add_comparisons(comparisons)

Now that all of the comparison information has been added, we next need to create the hierarchy and add it to the Compose object. A hierarchy is simply a dictionary in which the keys are the names of parent Compare objects and the values are lists of the names of their children:

>>> hierarchy = {'Criteria': ['Cost', 'Safety', 'Style', 'Capacity'],
		 'Cost': ['Price', 'Fuel', 'Resale', 'Maintenance'],
		 'Capacity': ['Passenger', 'Cargo']}

>>> compose.add_hierarchy(hierarchy)

With these two steps complete, we can now view the synthesized results of the analysis.

We view a report for a Compose object in the same way we do for a Compare object. The only difference is that the Compose object displays a complete report by default; in order to view the report of a single Compare object in the hierarchy, we need to specify its name:

>>> criteria_report = compose.report('Criteria', show=True)
{
    "name": "Criteria",
    "global_weight": 1.0,
    "local_weight": 1.0,
    "target_weights": {
        "Odyssey": 0.218,
        "Accord Sedan": 0.21,
        "Element": 0.161,
        "Accord Hybrid": 0.154,
        "CR-V": 0.149,
        "Pilot": 0.108
    },
    "elements": {
        "global_weights": {
            "Cost": 0.51,
            "Safety": 0.234,
            "Capacity": 0.215,
            "Style": 0.041
        },
        "local_weights": {
            "Cost": 0.51,
            "Safety": 0.234,
            "Capacity": 0.215,
            "Style": 0.041
        },
        "consistency_ratio": 0.08
    }
}

We can access the public properties of the comparison information we've added to the Compose object using either dot or bracket notation:

>>> print(compose.Criteria.target_weights)
{'Odyssey': 0.218, 'Accord Sedan': 0.21, 'Element': 0.161, 'Accord Hybrid': 0.154, 'CR-V': 0.149, 'Pilot': 0.108}

>>> print(compose['Resale']['local_weights'])
{'CR-V': 0.188, 'Accord Sedan': 0.177, 'Element': 0.164, 'Odyssey': 0.164, 'Accord Hybrid': 0.157, 'Pilot': 0.15}

We can see that normalizing the numeric criteria leads to a slightly different set of target weights, though the Odyssey and the Accord Sedan remain the top two vehicles to consider for purchase.

Details

Keep reading to learn the details of the AHPy library's API...

The Compare Class

The Compare class computes the weights and consistency ratio of a positive reciprocal matrix, created using an input dictionary of pairwise comparison values. Optimal values are computed for any missing pairwise comparisons. Compare objects can also be linked together to form a hierarchy representing the decision problem: the target weights of the problem elements are then derived by synthesizing all levels of the hierarchy.

Compare(name, comparisons, precision=4, random_index='dd', iterations=100, tolerance=0.0001, cr=True)

name: str (required), the name of the Compare object

  • This property is used to link a child object to its parent and must be unique

comparisons: dict (required), the elements and values to be compared, provided in one of two forms:

  1. A dictionary of pairwise comparisons, in which each key is a tuple of two elements and each value is their pairwise comparison value

    • {('a', 'b'): 3, ('b', 'c'): 2, ('a', 'c'): 5}
    • The order of the elements in the key matters: the comparison ('a', 'b'): 3 means "a is moderately more important than b"
  2. A dictionary of measured values, in which each key is a single element and each value is that element's measured value

    • {'a': 1.2, 'b': 2.3, 'c': 3.4}
    • Given this form, AHPy will automatically create consistent, normalized target weights

precision: int, the number of decimal places to take into account when computing both the target weights and the consistency ratio of the Compare object

  • The default precision value is 4

random_index: 'dd' or 'saaty', the set of random index estimates used to compute the Compare object's consistency ratio

  • 'dd' supports the computation of consistency ratios for matrices less than or equal to 100 × 100 in size and uses estimates from:

    Donegan, H.A. and Dodd, F.J., 'A Note on Saaty's Random Indexes,' Mathematical and Computer Modelling, 15:10, 1991, pp. 135-137 (DOI: 10.1016/0895-7177(91)90098-R)

  • 'saaty' supports the computation of consistency ratios for matrices less than or equal to 15 × 15 in size and uses estimates from:

    Saaty, T., Theory And Applications Of The Analytic Network Process, Pittsburgh: RWS Publications, 2005, p. 31

  • The default random index is 'dd'

iterations: int, the stopping criterion for the algorithm used to compute the Compare object's target weights

  • If target weights have not been determined after this number of iterations, the algorithm stops and the last principal eigenvector to be computed is used as the target weights
  • The default number of iterations is 100

tolerance: float, the stopping criterion for the cycling coordinates algorithm used to compute the optimal value of missing pairwise comparisons

  • The algorithm stops when the difference between the norms of two cycles of coordinates is less than this value
  • The default tolerance value is 0.0001

cr: bool, whether to compute the target weights' consistency ratio

  • Set cr=False to compute the target weights of a matrix when a consistency ratio cannot be determined due to the size of the matrix
  • The default value is True

The properties used to initialize the Compare class are intended to be accessed directly, along with a few others:

Compare.global_weight: float, the global weight of the Compare object within the hierarchy

Compare.local_weight: float, the local weight of the Compare object within the hierarchy

Compare.global_weights: dict, the global weights of the Compare object's elements; each key is an element and each value is that element's computed global weight

  • {'a': 0.25, 'b': 0.25}

Compare.local_weights: dict, the local weights of the Compare object's elements; each key is an element and each value is that element's computed local weight

  • {'a': 0.5, 'b': 0.5}

Compare.target_weights: dict, the target weights of the elements in the lowest level of the hierarchy; each key is an element and each value is that element's computed target weight; if the global weight of the Compare object is less than 1.0, the value will be None

  • {'a': 0.5, 'b': 0.5}

Compare.consistency_ratio: float, the consistency ratio of the Compare object's pairwise comparisons

Compare.add_children()

Compare objects can be linked together to form a hierarchy representing the decision problem. To link Compare objects together into a hierarchy, call add_children() on the Compare object intended to form the upper level (the parent) and include as an argument a list or tuple of one or more Compare objects intended to form its lower level (the children).

In order to properly synthesize the levels of the hierarchy, the name of each child object MUST appear as an element in its parent object's input comparisons dictionary.

Compare.add_children(children)

children: list or tuple (required), the Compare objects that will form the lower level of the current Compare object

>>> child1 = ahpy.Compare(name='child1', ...)
>>> child2 = ahpy.Compare(name='child2', ...)

>>> parent = ahpy.Compare(name='parent', comparisons={('child1', 'child2'): 5})
>>> parent.add_children([child1, child2])

The precision of the target weights is updated as the hierarchy is constructed: each time add_children() is called, the precision of the target weights is set to equal that of the Compare object with the lowest precision in the hierarchy. Because lower precision propagates up through the hierarchy, the target weights will always have the same level of precision as the hierarchy's least precise Compare object. This also means that it is possible for the precision of a Compare object's target weights to be different from the precision of its local and global weights.

Compare.report()

A standard report on the details of a Compare object is available. To return the report as a dictionary, call report() on the Compare object; to simultaneously print the information to the console in JSON format, set show=True. The report is available in two levels of detail; to return the most detailed report, set verbose=True.

Compare.report(complete=False, show=False, verbose=False)

complete: bool, whether to return a report for every Compare object in the hierarchy

  • This returns a dictionary of reports, with the keys of the dictionary being the names of the Compare objects
    • {'a': {'name': 'a', ...}, 'b': {'name': 'b', ...}}
  • The default value is False

show: bool, whether to print the report to the console in JSON format

  • The default value is False

verbose: bool, whether to include full details of the Compare object in the report

  • The default value is False

The keys of the report take the following form:

name: str, the name of the Compare object

global_weight: float, the global weight of the Compare object within the hierarchy

local_weight: float, the local weight of the Compare object within the hierarchy

target_weights: dict, the target weights of the elements in the lowest level of the hierarchy; each key is an element and each value is that element's computed target weight

  • {'a': 0.5, 'b': 0.5}
  • If the global weight of the Compare object is less than 1.0, the value will be None

elements: dict, information regarding the elements compared by the Compare object

  • global_weights: dict, the global weights of the Compare object's elements; each key is an element and each value is that element's computed global weight
    • {'a': 0.25, 'b': 0.25}
  • local_weights: dict, the local weights of the Compare object's elements; each key is an element and each value is that element's computed local weight
    • {'a': 0.5, 'b': 0.5}
  • consistency_ratio: float, the consistency ratio of the Compare object's pairwise comparisons

The remaining dictionary keys are only displayed when verbose=True:

  • random_index: 'Donegan & Dodd' or 'Saaty', the random index used to compute the consistency ratio
  • count: int, the number of elements compared by the Compare object
  • names: list, the names of the elements compared by the Compare object

children: dict, the children of the Compare object

  • count: int, the number of the Compare object's children
  • names: list, the names of the Compare object's children
  • If the Compare object has no children, the value will be None

comparisons: dict, the comparisons of the Compare object

  • count: int, the number of comparisons made by the Compare object, not counting reciprocal comparisons
  • input: dict, the comparisons input to the Compare object; this is identical to the input comparisons dictionary
  • computed: dict, the comparisons computed by the Compare object; each key is a tuple of two elements and each value is their computed pairwise comparison value
    • {('c', 'd'): 0.730297106886979}, ...}
  • If the Compare object has no computed comparisons, the value will be None

The Compose Class

The Compose class can store and structure all of the information making up a decision problem. After first adding comparison information to the object, then adding the problem hierarchy, the analysis results of the multiple different Compare objects can be accessed through the single Compose object.

Compose()

After adding all necessary information, the public properties of any stored Compare object can be accessed directly through the Compose object using either dot or bracket notation:

>>> my_compose_object.a.global_weights

>>> my_compose_object['a']['global_weights']

Compose.add_comparisons()

The comparison information of a decision problem can be added to a Compose object in any of the several ways listed below. Always add comparison information before adding the problem hierarchy.

Compose.add_comparisons(item, comparisons=None, precision=4, random_index='dd', iterations=100, tolerance=0.0001, cr=True)

item: Compare object, list or tuple, or string (required), this argument allows for multiple input types:

  1. A single Compare object

    • Compare('a', comparisons=a, ...)
  2. A list or tuple of Compare objects

    • [Compare('a', ...), Compare('b', ...)]
  3. The data necessary to create a Compare object

    • 'a', comparisons=a, precision=3, ...
    • The method signature mimics that of the Compare class for this reason
  4. A nested list or tuple of the data necessary to create a Compare object

    • (('a', a, 3, ...), ('b', b, 3, ...))

All other arguments are identical to those of the Compare class.

Compose.add_hierarchy()

The Compose class uses an abstract representation of the problem hierarchy to automatically link its Compare objects together. When a hierarchy is added, the elements of the decision problem are synthesized and the analysis results are immediately available for use or viewing.

Compose.add_hierarchy() should only be called AFTER all comparison information has been added to the Compose object.

Compose.add_hierarchy(hierarchy)

hierarchy: dict, a representation of the hierarchy as a dictionary, in which the keys are the names of parent Compare objects and the values are lists of the names of their children

  • {'a': ['b', 'c'], 'b': ['d', 'e']}

Compose.report()

The standard report available for a Compare object can be accessed through the Compose object. Calling report() on a Compose object is equivalent to calling report(complete=True) on a Compare object and will return a dictionary of all the reports within the hierarchy; calling report(name='a') on a Compose object is equivalent to calling a.report() on the named Compare object.

Compose.report(name=None, show=False, verbose=False)

name: str, the name of a Compare object report to return; if None, returns a dictionary of reports, with the keys of the dictionary being the names of the Compare objects in the hierarchy

  • {'a': {'name': 'a', ...}, 'b': {'name': 'b', ...}}
  • The default value is None

All other arguments are identical to the Compare class's report() method.

A Note on Weights

Compare objects compute up to three kinds of weights for their elements: global weights, local weights and target weights. Compare objects also compute their own global and local weight, given their parent.

  • Global weights display the computed weights of a Compare object's elements dependent on that object's global weight within the current hierarchy

    • Global weights are derived by multiplying the local weights of the elements within a Compare object by that object's own global weight in the current hierarchy
  • Local weights display the computed weights of a Compare object's elements independent of that object's global weight within the current hierarchy

    • The local weights of the elements within a Compare object will always (approximately) sum to 1.0
  • Target weights display the synthesized weights of the problem elements described in the lowest level of the current hierarchy

    • Target weights are only available from the Compare object at the highest level of the hierarchy (i.e. the only Compare object without a parent)

N.B.

A Compare object that does not have a parent will have identical global and local weights; a Compare object that has neither a parent nor children will have identical global, local and target weights.

In many instances, the sum of the local or target weights of a Compare object will not equal 1.0 exactly. This is due to rounding. If it's critical that the sum of the weights equals 1.0, it's recommended to simply divide the weights by their cumulative sum: x = x / np.sum(x). Note, however, that the resulting values will contain a false level of precision, given their inputs.

Missing Pairwise Comparisons

When a Compare object is initialized, the elements forming the keys of the input comparisons dictionary are permuted. Permutations of elements that do not contain a value within the input comparisons dictionary are then optimally solved for using the cyclic coordinates algorithm described in:

Bozóki, S., Fülöp, J. and Rónyai, L., 'On optimal completion of incomplete pairwise comparison matrices,' Mathematical and Computer Modelling, 52:1–2, 2010, pp. 318-333 (DOI: 10.1016/j.mcm.2010.02.047)

As the paper notes, "The number of necessary pairwise comparisons ... depends on the characteristics of the real decision problem and provides an exciting topic of future research" (29). In other words, don't rely on the algorithm to fill in a comparison dictionary that has a large number of missing values: it certainly might, but it also very well might not. Caveat emptor!

The example below demonstrates this functionality of AHPy using the following matrix:

a b c d
a 1 1 5 2
b 1 1 3 4
c 1/5 1/3 1 3/4
d 1/2 1/4 4/3 1

We'll first compute the target weights and consistency ratio for the complete matrix, then repeat the process after removing the (c, d) comparison marked in bold. We can view the computed value in the Compare object's detailed report:

>>> comparisons = {('a', 'b'): 1, ('a', 'c'): 5, ('a', 'd'): 2,
		   ('b', 'c'): 3, ('b', 'd'): 4,
		   ('c', 'd'): 3 / 4}

>>> complete = ahpy.Compare('Complete', comparisons)

>>> print(complete.target_weights)
{'b': 0.3917, 'a': 0.3742, 'd': 0.1349, 'c': 0.0991}

>>> print(complete.consistency_ratio)
0.0372

>>> del comparisons[('c', 'd')]

>>> missing_cd = ahpy.Compare('Missing_CD', comparisons)

>>> print(missing_cd.target_weights)
{'b': 0.392, 'a': 0.3738, 'd': 0.1357, 'c': 0.0985}

>>> print(missing_cd.consistency_ratio)
0.0372

>>> report = missing_cd.report(verbose=True)
>>> print(report['comparisons']['computed'])
{('c', 'd'): 0.7302971068355002}

Development and Testing

To set up a development environment and run the included tests, you can use the following commands:

virtualenv .venv
source .venv/bin/activate
python setup.py develop
pip install pytest
pytest