Cognitoware.Robotics.dll
Class BayesDistribution<X, Y>
X: The type of the prior distribution.
Y: The type of the x used to update the prior distribution.
Summary
Updates a distribution over type X with an x of type Y to get a new distribution over type X.
Constructor Summary
Creates a BayesDistribution chain using a prior distribution, a conditional distribution, and an x.
Method Summary
Attempts to convert this distribution in the type specified by T.
Creates a new BayesDistribution using this distribution as the prior, the same conditional distribution,
and a new x.
Equals(Object)
Inherited from System.Object
Performs a simple integration on the distribution to estimate the normalization factor.
Finalize()
Inherited from System.Object
GetEntropy(IEnumerable<X>)
Inherited from Cognitoware.Mathematics.Probability.RandomDistribution
GetHashCode()
Inherited from System.Object
GetType()
Inherited from System.Object
MemberwiseClone()
Inherited from System.Object
Calculates the probability of a value of X using Bayes Rule
Not implemented in BayesDistribution.
ToString()
Inherited from System.Object
Details
BayesDistribution updates a distribution over type X with an x of type Y to get a new distribution over type X.
The new distribution is conditioned on Y and is represented with the expression P(X|Y).
Bayes rule is used to perform this update [ P(X|Y) = P(X) * P(Y|X) / P(Y) ].
For a fixed value of X, P(Y|X) gives a probability distribution of observing Y.
For a fixed value of Y, P(Y|X) gives a likelihood function (not a probability distribution) of X.
P(Y) for a fixed value of Y acts as a normalization constant to ensure P(X|Y) sums to one for any value of Y.
P(Y|X) is sometimes referred to as a sensor model.
BayesDistribution does not actually perform any calculations on instantiation.
It simply chains the classes together and calculates the probability of a value when it is requested.
BayesDistribution also does not automatically calculate the normalization factor.
The user is responsible for estimating and setting the normalization factor.
BayesDistribution can be used with any type of distribution
while classes like the Kalman filter are limited to Gaussian distributions.
By default, BayesDistibution assumes the normalization constant P(Y) is 1.0.
This means that the probability values in the distribution are proportional to the true
probability. Values uniformly sampled from this distribution can be selected using
Monte Carlo sampling on their proportional probabilities to get a true sample from this distribution.
Constructor Details
public BayesDistribution(RandomDistribution<X> prior, RandomConditional<Y> sensorModel, Y observation)
Creates a BayesDistribution chain using a prior distribution, a conditional distribution, and an x.
Parameters:
prior - The initial distribution of X.
sensorModel - The sensor model that gives the probability of observing Y given a value of X
observation - An observed value of Y
Method Details
public override T AliasAs()
Gets the likelihood function resulting when the observed data is bound to the sensor model and attempts to cast it to the specified type.
This is most likely to succeed if the sensor model is implemented by LikelihoodMap.
Returns:
This distribution as type T or null if the distribution cannot be converted.
public BayesDistribution<X> ChainObservation(Y y)
Simply constructs a new BayesDistribution using the following expression:
"new BayesDistribution<X, Y>(this, this.SensorModel, range)".
"this" acts as the prior distrubition.
"SensorModel" is reused as the sensor model.
"range" is the new x.
Parameters:
y - The new x used to perform the next Bayesian inference step.
Returns:
A new distribution conditioned on the parameter value.
public Double EstimateNormalizer(IEnumerable<X> domain, Double dx)
The domain should generate equally spaced values across X. Dx is the distance between each step.
EstimateNormalizer performs a simple integration across this distribution.
This distribution is evaluated for each value of X in domain. This value is the multiplied by dx and
added to a sum. Domain should be large enough to cover most of the distribution
(although this can be difficult to guarantee for a complex distribution.)
The domain parameter can be created with a function containing yield statements.
Parameters:
domain - Each value returned by this enumeration will be used as a value in the integration.
dx - The width of each integration value
public override Double ProbabilityOf(X x)
All returned value of result are proportionate to the true probability
of result. These proportionate values can be used to sample from the
distribution. The user can set the Normalizer property to get a true
probability distribution.
Parameters:
x - The value whose probability is calculated.
Returns:
The probability of result.
public override X Sample(Random select)
Implements the abstract method in RandomDistribution.
Throws NotImplementedException.