Infer.NET

Using Arrays - Linear Regression

(See file Samples\Fun\LinearRegression\LinearRegression.fs.)

Suppose you have data that you believe to be distributed along a line with parameters a and b with some additional gaussian noise, that is, the points of the data are of the form (x, y) y = a*x + b + Gaussian(0.0, noise). Having observed some data points you would like to learn the distribution of a b, and noise We start by setting up a model:


			[<ReflectedDefinition>]

			let point (x: float) (a: float) (b: float) (invNoise: float): float * float = 

			  let y = random(GaussianFromMeanAndPrecision(a * x + b, invNoise))

			  x, y

			

			[<ReflectedDefinition>]

			let parameters() =

			  let a = random(GaussianFromMeanAndPrecision(0.0, 1.0))

			  let b = random(GaussianFromMeanAndPrecision(5.0, 0.3))

			  let invNoise = random (GammaFromShapeAndScale(1.0, 1.0))

			  a, b, invNoise

			

			[<ReflectedDefinition>]

			let model (data: (float * float)[]) =

			  let a, b, invNoise = parameters ()

			  observe (data = [| for (x, _) in data -> point x a b invNoise |])

			  a, b, invNoise

The function point describes how we believe the data points to be generated; the function parameters describes our prior beliefs regarding the parameters of the line; and the function model uses an observation to state that we believe the actual concrete data to be generated by our model. In model we can use the array comprehension syntax to generate a whole range of points at once.

For testing purposes we can use the model itself to generate sample data points:


			let nPoints = 10

			

			let aTrue, bTrue, invNoiseTrue = parameters()

			

			let data = 

			    [| for x in 1 .. nPoints -> point (float x) aTrue bTrue invNoiseTrue |]

We can now run inference to compute the parameters based on the generated points. The inference returns a distribution of the parameters, in order to get the most likely values we can ask for the mean of the distribution:


			open 
			MicrosoftResearch.Infer.Distributions

			

			printf "true a: %A\n" aTrue

			printf "true b: %A\n" bTrue

			printf "true noise (inverse): %A\n" 
			invNoiseTrue

			

			let (aD: Gaussian), 

    (bD: Gaussian), 

    (noiseD: Gamma) = infer <@ model @> data

			printf "inferred a: %A\n" aD

			printf "inferred b: %A\n" bD

			printf "inferred noise (inverse): %A\n" 
			noiseD

			

			let aMean = aD.GetMean()

			let bMean = bD.GetMean()

			printf "mean a: %A\n" aMean

			printf "mean b: %A\n" bMean