Added Trend Parameter encoding in GP Priors

Author: mandar2812

build.sbt

@@ -12,7 +12,7 @@ packageDescription := "DynaML is a scala library/repl for implementing and worki
   "which can be extended easily to implement advanced models for small and large scale applications.\n\n"+
   "But the library can also be used as an educational/research tool for data analysis."
 
-val mainVersion = "v1.4.3-beta.1"
+val mainVersion = "v1.4.3-beta.2"
 
 val dataDirectory = settingKey[File]("The directory holding the data files for running example scripts")
 

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/bayes/GaussianProcessPrior.scala

@@ -28,7 +28,7 @@ import io.github.mandar2812.dynaml.modelpipe.GPRegressionPipe2
 import io.github.mandar2812.dynaml.models.gp.AbstractGPRegressionModel
 
 import scala.reflect.ClassTag
-import io.github.mandar2812.dynaml.pipes.{DataPipe, MetaPipe, ParallelPipe}
+import io.github.mandar2812.dynaml.pipes.{DataPipe, Encoder, MetaPipe, ParallelPipe}
 import io.github.mandar2812.dynaml.probability.{MatrixNormalRV, MultGaussianPRV}
 
 /**
@@ -46,7 +46,8 @@ import io.github.mandar2812.dynaml.probability.{MatrixNormalRV, MultGaussianPRV}
   * */
 abstract class GaussianProcessPrior[I: ClassTag, MeanFuncParams](
   val covariance: LocalScalarKernel[I],
-  val noiseCovariance: LocalScalarKernel[I]) extends
+  val noiseCovariance: LocalScalarKernel[I],
+  val trendParamsEncoder: Encoder[MeanFuncParams, Map[String, Double]]) extends
   StochasticProcessPrior[
     I, Double, PartitionedVector,
     MultGaussianPRV, MultGaussianPRV,
@@ -138,6 +139,7 @@ abstract class GaussianProcessPrior[I: ClassTag, MeanFuncParams](
     ScaledKernel[I](self.covariance, scalingFunc),
     ScaledKernel[I](self.noiseCovariance, scalingFunc),
     MetaPipe((p: MeanFuncParams) => (x: I) => self.meanFunctionPipe(p)(x)*scalingFunc(x)),
+    self.trendParamsEncoder,
     self._meanFuncParams)
 
 }
@@ -151,8 +153,9 @@ object GaussianProcessPrior {
     covariance: LocalScalarKernel[I],
     noiseCovariance: LocalScalarKernel[I],
     meanFPipe: MetaPipe[MeanFuncParams, I, Double],
+    trendParamsEncoder: Encoder[MeanFuncParams, Map[String, Double]],
     initialParams: MeanFuncParams): GaussianProcessPrior[I, MeanFuncParams] =
-    new GaussianProcessPrior[I, MeanFuncParams](covariance, noiseCovariance) {
+    new GaussianProcessPrior[I, MeanFuncParams](covariance, noiseCovariance, trendParamsEncoder) {
 
       private var params = initialParams
 
@@ -173,9 +176,10 @@ object GaussianProcessPrior {
   * */
 class LinearTrendGaussianPrior[I: ClassTag](
   cov: LocalScalarKernel[I], n: LocalScalarKernel[I],
+  trendParamsEncoder: Encoder[(I, Double), Map[String, Double]],
   trendParams: I, intercept: Double)(
   implicit inner: InnerProductSpace[I, Double]) extends
-  GaussianProcessPrior[I, (I, Double)](cov, n) with
+  GaussianProcessPrior[I, (I, Double)](cov, n, trendParamsEncoder) with
   LinearTrendStochasticPrior[I, MultGaussianPRV, MultGaussianPRV, AbstractGPRegressionModel[Seq[(I, Double)], I]]{
 
   override val innerProduct = inner
@@ -210,8 +214,12 @@ class LinearTrendGaussianPrior[I: ClassTag](
   * */
 abstract class CoRegGPPrior[I: ClassTag, J: ClassTag, MeanFuncParams](
   covarianceI: LocalScalarKernel[I], covarianceJ: LocalScalarKernel[J],
-  noiseCovarianceI: LocalScalarKernel[I], noiseCovarianceJ: LocalScalarKernel[J]) extends
-  GaussianProcessPrior[(I,J), MeanFuncParams](covarianceI:*covarianceJ, noiseCovarianceI:*noiseCovarianceJ) {
+  noiseCovarianceI: LocalScalarKernel[I], noiseCovarianceJ: LocalScalarKernel[J],
+  trendParamsEncoder: Encoder[MeanFuncParams, Map[String, Double]]) extends
+  GaussianProcessPrior[(I,J), MeanFuncParams](
+    covarianceI:*covarianceJ,
+    noiseCovarianceI:*noiseCovarianceJ,
+    trendParamsEncoder) {
 
   self =>
 
@@ -243,6 +251,7 @@ abstract class CoRegGPPrior[I: ClassTag, J: ClassTag, MeanFuncParams](
       MetaPipe((p: MeanFuncParams) => (x: (I, J)) => {
         self.meanFunctionPipe(p)(x)*scalingFunc._1(x._1)*scalingFunc._2(x._2)
       }),
+      self.trendParamsEncoder,
       self._meanFuncParams
     )
 }
@@ -268,8 +277,12 @@ object CoRegGPPrior {
     covarianceI: LocalScalarKernel[I], covarianceJ: LocalScalarKernel[J],
     noiseCovarianceI: LocalScalarKernel[I], noiseCovarianceJ: LocalScalarKernel[J])(
     meanFPipe: MetaPipe[MeanFuncParams, (I, J), Double],
+    trendParamsEncoder: Encoder[MeanFuncParams, Map[String, Double]],
     initialParams: MeanFuncParams) =
-    new CoRegGPPrior[I, J, MeanFuncParams](covarianceI, covarianceJ, noiseCovarianceI, noiseCovarianceJ) {
+    new CoRegGPPrior[I, J, MeanFuncParams](
+      covarianceI, covarianceJ,
+      noiseCovarianceI, noiseCovarianceJ,
+      trendParamsEncoder) {
 
       private var params = initialParams
 

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/bayes/StochasticProcessPrior.scala

@@ -27,14 +27,15 @@ import spire.algebra.InnerProductSpace
 
 
 /**
-  * @author mandar2812 date: 16/02/2017.
-  *
   * Represents a stochastic process with a
   * defined prior distribution.
-  */
+  *
+  * @author mandar2812 date: 16/02/2017.
+  *
+  * */
 trait StochasticProcessPrior[I, Y, Y1,
 D <: RandomVarWithDistr[Y1, _], W,
-StochasticModel <: StochasticProcessModel[Seq[(I, Y)], I, Y, W]] {
+StochasticModel <: StochasticProcessModel[Seq[(I, Y)], I, Y, W]] extends Serializable {
 
   protected var hyperPrior: Map[String, ContinuousDistrRV[Double]] = Map()
 

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/probability/RandomVariable.scala

@@ -37,7 +37,7 @@ import io.github.mandar2812.dynaml.utils._
   *                the random variable is defined;
   *                i.e. the range of values it can take
   */
-abstract class RandomVariable[Domain] {
+abstract class RandomVariable[Domain] extends Serializable {
 
   self =>
   /**
@@ -91,7 +91,7 @@ abstract class RandomVariable[Domain] {
   * underlying probability density function.
   *
   * */
-trait HasDistribution[Domain] {
+trait HasDistribution[Domain] extends Serializable {
 
   /**
     * The actual probability density

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/probability/RejectionSamplingScheme.scala

@@ -36,7 +36,7 @@ JointDist <: Density[(ConditioningSet, Domain)] with Rand[(ConditioningSet, Doma
 
   override val sample = prior.sample >
     DataPipe[ConditioningSet, ConditioningSet, Domain](
-      (c: ConditioningSet) => (c, likelihood(c).sample())
+      (c: ConditioningSet) => (c, likelihood(c).draw)
     )
 
   override val posterior: DataPipe[Domain, RandomVariable[ConditioningSet]] =

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/probability/distributions/GenericDistribution.scala

@@ -5,7 +5,7 @@ import breeze.stats.distributions.{ContinuousDistr, Density, DiscreteDistr, Rand
 /**
   * @author mandar2812 date 06/01/2017.
   * */
-abstract class GenericDistribution[T] extends Density[T] with Rand[T]
+abstract class GenericDistribution[T] extends Density[T] with Rand[T] with Serializable
 
 abstract class AbstractDiscreteDistr[T] extends GenericDistribution[T] with DiscreteDistr[T]
 
@@ -15,7 +15,7 @@ abstract class AbstractContinuousDistr[T] extends GenericDistribution[T] with Co
   * Distributions which can generate confidence intervals around their mean
   * can extend this trait and override the `confidenceInterval` method .
   * */
-trait HasErrorBars[T] {
+trait HasErrorBars[T] extends Serializable {
 
   def confidenceInterval(s: Double): (T, T)
 }

scripts/probModels.sc

@@ -45,7 +45,7 @@ val abcPosterior = abcModel.posterior(data)
 
 val posterior: RandomVariable[(Double, Double)] = gModel.posterior(data)
 
-val samples = (1 to 3000).map(_ => {posterior.sample()})
+val samples = (1 to 3000).map(_ => posterior.draw)
 
 histogram(data)
 title("Histogram of data")
@@ -73,7 +73,7 @@ c_model.tolerance_(0.05)
 
 val post = c_model.posterior(350)
 
-val postSamples = (1 to 2000).map(_ => post.sample())
+val postSamples = (1 to 2000).map(_ => post.draw)
 
 histogram((1 to 2000).map(_ => p.sample()))
 hold()

scripts/stochasticPriors.sc

@@ -13,6 +13,11 @@ val encoder = Encoder(
   (conf: Map[String, Double]) => (conf("c"), conf("s")),
   (cs: (Double, Double)) => Map("c" -> cs._1, "s" -> cs._2))
 
+val trendEncoder = Encoder(
+  (cs: (Double, Double)) => Map("slope" -> cs._1, "intercept" -> cs._2),
+  (conf: Map[String, Double]) => (conf("slope"), conf("intercept"))
+)
+
 val hyp_prior: Map[String, ContinuousDistrRV[Double]] = Map(
   "c" -> GaussianRV(2.5, 1.5),
   "s" -> RandomVariable(Gamma(2.0, 2.0)),
@@ -23,7 +28,7 @@ val sgp_hyp_prior = hyp_prior ++ Map("cutoff" -> GaussianRV(0.0, 1.0), "skewness
 val gsmKernel = GaussianSpectralKernel[Double](3.5, 2.0, encoder)
 val n = new MAKernel(0.8)
 
-val gp_prior = new LinearTrendGaussianPrior[Double](gsmKernel, n, 0.0, 0.0)
+val gp_prior = new LinearTrendGaussianPrior[Double](gsmKernel, n, trendEncoder, 0.0, 0.0)
 //gp_prior.hyperPrior_(hyp_prior)
 
 val sgp_prior = new LinearTrendESGPrior[Double](gsmKernel, n, 0.75, 0.1, 0.0, 0.0)