Package microsim.alignment.multiple

Class LogitScalingAlignment<T>

java.lang.Object

microsim.alignment.multiple.LogitScalingAlignment<T>

Type Parameters:

T - The type parameter usually representing the agent class.

All Implemented Interfaces:

AlignmentUtils<T>

public class LogitScalingAlignment<T>
extends Object
implements AlignmentUtils<T>

Multinomial alignment methods, where there is in general a set 'A' (>=2) of possible outcomes/states to align. This algorithm is called Logit Scaling and is based on the minimization of the information loss (relative entropy). This class can be applied to both weighted and non-weighted collections of agents. Additionally, it can be used in conventional binary alignment problems.

See Also:

• Peter Stephensen, A General Method for Alignment in Microsimulation models, International Journal of Microsimulation (2016) 9(3) 89-102

Implementation Note:

The design of this class suboptimal at the moment to make all class implementations share as much code as possible. In particular, when agents have no weight, they all automatically get one that is equal to 1. This code extends the original algorithm by adding the case of weighted samples. It also does add a filter for the whole population to align only the relevant subpopulation.

In the degenerate case when all probabilities are given as open-point distributions this algorithm fails to converge. This happens due to the fact it conserves 0s and 1s, thus, the entire array of probabilities does not change.

Field Summary

Fields

Modifier and Type	Field	Description
static final double	ERROR_THRESHOLD	Constant: 1.0E-15.

Constructor Summary

Constructors

Constructor	Description
LogitScalingAlignment()

Method Summary

Modifier and Type	Method	Description
final void	align(@NonNull Collection<T> agents, @Nullable org.apache.commons.collections4.Predicate<T> filter, @NonNull AlignmentMultiProbabilityClosure<T> closure, double @NonNull [] targetProbabilityDistribution)	General alignment procedure, it adjusts probabilities using all the provided parameters until the algorithm reaches the target precision/number of iterations.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface microsim.alignment.AlignmentUtils

extractAgentList, validateProbabilityValue

Field Details

ERROR_THRESHOLD

public static final double ERROR_THRESHOLD

Constant: 1.0E-15.

See Also:

• Constant Field Values

Constructor Details

LogitScalingAlignment

public LogitScalingAlignment()

Method Details

align

public final void align(@NonNull
 @NonNull Collection<T> agents,
 @Nullable
 @Nullable org.apache.commons.collections4.Predicate<T> filter,
 @NonNull
 @NonNull AlignmentMultiProbabilityClosure<T> closure,
 double @NonNull [] targetProbabilityDistribution)

General alignment procedure, it adjusts probabilities using all the provided parameters until the algorithm reaches the target precision/number of iterations.

Parameters:

agents - A collection of agents of a given type. Most commonly, every agent is a person, however, this class does not limit its usage to humans only.

filter - A filter to select a subpopulation from a given collection of agents, can accept null values.

closure - An object that specifies how to get the (unaligned) probability of the agent and how to set the result of the aligned probability.

targetProbabilityDistribution - The target discrete probability distribution. Means of the aligned probabilities must be equal to these values.

Throws:

NullPointerException - when agents, closure, or targetProbabilityDistribution, or any combination of them are null.

Implementation Requirements:

When the list of agents is empty, no alignment happens. The same works for cases when filtering produces an empty list.

Implementation Note:

The total number of iterations defaults to 50. The error threshold of the numerical scheme is 1e-15.