Skip to content

A Scala library for Bayesian Inference and Probabilistic Programming

License

Notifications You must be signed in to change notification settings

ruippeixotog/think-bayes-scala

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Think Bayes in Scala

A Scala implementation of the classes and functions used in the great book Think Bayes by Allen B. Downey, available for free (and open-source) here.

Quick start

The code in this repository is available as a library and can be used in Scala 2.11.x projects by adding the following dependency to build.sbt:

libraryDependencies += "net.ruippeixotog" %% "think-bayes" % "0.1"

Core classes

Probability mass functions

The Pmf class is arguably the core collection in Think Bayes, due to the latter's focus on problem solving using discrete approximations instead of continuous mathematics. The way to build a Pmf and manipulate it is pretty simple:

  scala> import thinkbayes._
  import thinkbayes._

  scala> val pmf = Pmf('a' -> 0.2, 'b' -> 0.2, 'c' -> 0.6)
  pmf: thinkbayes.Pmf[Char] = Map(a -> 0.2, b -> 0.2, c -> 0.6)

  scala> pmf.prob('a')
  res0: Double = 0.2

  scala> pmf.prob(_ < 'c')
  res1: Double = 0.4

  scala> pmf.sample()
  res2: Char = c

  scala> pmf.printChart()
  a 0.2    ##########
  b 0.2    ##########
  c 0.6    ##############################

A Pmf is implemented as an immutable map and can be used as such:

  scala> pmf.size
  res3: Int = 3

  scala> pmf.map { case (k, v) => ((k + 1).toChar, v) }
  res4: thinkbayes.Pmf[Char] = Map(b -> 0.2, c -> 0.2, d -> 0.6)

  scala> pmf.filter(_._1 == 'a').normalized
  res5: thinkbayes.Pmf[Char] = Map(a -> 1.0)

  scala> pmf.foldLeft("")(_ + _._1)
  res6: String = abc

  scala> pmf.toList
  res7: List[(Char, Double)] = List((a,0.2), (b,0.2), (c,0.6))

Specialized Pmf merging methods can model more complex problems in a very concise manner:

  scala> def die(n: Int) = Pmf(1 to n)
  die: (n: Int)thinkbayes.Pmf[Int]

  scala> die(6)
  res8: thinkbayes.Pmf[Int] = Map(5 -> 0.16666666666666666, 1 -> 0.16666666666666666, 6 -> 0.16666666666666666, 2 -> 0.16666666666666666, 3 -> 0.16666666666666666, 4 -> 0.16666666666666666)

  scala> die(6).mean
  res9: Double = 3.5

  scala> (die(6) ++ die(6)).printChart() // sum of two dice
  2  0.0277 #
  3  0.0555 ##
  4  0.0833 ####
  5  0.1111 #####
  6  0.1388 ######
  7  0.1666 ########
  8  0.1388 ######
  9  0.1111 #####
  10 0.0833 ####
  11 0.0555 ##
  12 0.0277 #

  scala> val bag = Pmf(List(die(4), die(6), die(8), die(12), die(20))) // a bag containing 5 different dice
  bag: thinkbayes.Pmf[thinkbayes.Pmf[Int]] = Map(Map(5 -> 0.08333333333333333, 10 -> 0.08333333333333333, 1 -> 0.08333333333333333, 6 -> 0.08333333333333333, 9 -> ...

  scala> bag.mixture.printChart() // roll of a random die from the bag
  1  0.135  ######
  2  0.135  ######
  3  0.135  ######
  4  0.135  ######
  5  0.0850 ####
  6  0.0850 ####
  7  0.0516 ##
  8  0.0516 ##
  9  0.0266 #
  10 0.0266 #
  11 0.0266 #
  12 0.0266 #
  13 0.0100
  14 0.0100
  15 0.0100
  16 0.0100
  17 0.0100
  18 0.0100
  19 0.0100
  20 0.0100

The Distributions extension provides methods for creating common Pmf such as Gaussian or Poisson distributions.

Bayesian suites

The implementation of Suite provided in this library does not extend Pmf; it is rather provided as a trait which applications can implement to model specific problems:

  scala> case class Dice(hypos: Seq[Int]) extends SimpleSuite[Int, Int] {
       |   val pmf = Pmf(hypos) // which dice from `hypos` are we rolling?
       |   def likelihood(data: Int, hypo: Int) = if(hypo < data) 0 else 1.0 / hypo
       | }
  defined class Dice

  scala> val prior = Dice(List(4, 6, 8, 12, 20))
  prior: Dice = Dice(List(4, 6, 8, 12, 20))

  scala> prior.printChart()
  4  0.2    ##########
  6  0.2    ##########
  8  0.2    ##########
  12 0.2    ##########
  20 0.2    ##########

  scala> val posterior = prior.observed(6) // after a 6 is rolled
  posterior: thinkbayes.Suite[Int,Int] = thinkbayes.Suite$$anon$1@120fb03e

  scala> posterior.printChart()
  4  0.0
  6  0.3921 ###################
  8  0.2941 ##############
  12 0.1960 #########
  20 0.1176 #####

The same prior could be built directly with:

  scala> val prior = Suite[Int, Int](Pmf(List(4, 6, 8, 12, 20))) { (d, h) =>
       |   if (h < d) 0 else 1.0 / h
       | }
  prior: thinkbayes.Suite[Int,Int]{val pmf: thinkbayes.Pmf[Int]} = thinkbayes.Suite$$anon$1@130dd39f

Multiple observations can be given to the Suite in bulk, which can yield results more stable numerically:

  scala> posterior.observed(6, 8, 7, 7, 5, 4).printChart()
  4  0.0
  6  0.0
  8  0.9432 ###############################################
  12 0.0552 ##
  20 0.0015

Cumulative distribution functions

A Cdf can be created just like a Pmf. It supports efficient querying for the cumulative probability on a given value (prob) and for the value at a given percentile (value):

  scala> val cdf = Cdf('a' -> 0.2, 'b' -> 0.2, 'c' -> 0.6)
  cdf: thinkbayes.Cdf[Char] = CategoricalCdf(Vector((a,0.2), (b,0.4), (c,1.0)))

  scala> cdf.prob('b')
  res10: Double = 0.4

  scala> cdf.value(0.5)
  res11: Char = c

  scala> cdf.value(0.35)
  res12: Char = b

  scala> cdf.printChart()
  a 0.2    ##########
  b 0.4    ####################
  c 1.0    ##################################################

Unlike Pmf, Cdf does not implement the Map trait and, therefore, does not inherit the common Scala collection methods. If you need to use those, you can convert easily a Cdf to and from a Pmf:

  scala> cdf.toPmf
  res13: thinkbayes.Pmf[Char] = Map(a -> 0.2, b -> 0.2, c -> 0.6)

  scala> cdf.toPmf.toCdf
  res14: thinkbayes.Cdf[Char] = CategoricalCdf(Vector((a,0.2), (b,0.4), (c,1.0)))

Probability density functions

A Pdf can be created from a Scala real-valued function and provides a density method for calculating the density at a given value:

  scala> val pdf = Pdf { x => math.max(-x * x + 1, 0) }
  pdf: thinkbayes.Pdf = thinkbayes.Pdf$$anon$3@744cb6e3

  scala> pdf.density(0)
  res15: Double = 1.0

  scala> pdf.density(0.5)
  res16: Double = 0.75

A BoundedPdf is a Pdf whose domain has known lower and upper bounds.

  scala> val bpdf = Pdf(-1.0, 1.0) { x => math.max(-x * x + 1, 0) }
  bpdf: thinkbayes.BoundedPdf{val lowerBound: Double; val upperBound: Double} = thinkbayes.Pdf$$anon$2@397820d5

Both can be converted to a Pmf given a range or sequence of discrete values to compute. A BoundedPdf can alternatively be given a step value only. In both cases, the probabilities of the returned Pmf are normalized:

  scala> pdf.toPmf(0.0 to 1.0 by 0.1).printChart()
  0.0                 0.1398 ######
  0.1                 0.1384 ######
  0.2                 0.1342 ######
  0.30000000000000004 0.1272 ######
  0.4                 0.1174 #####
  0.5                 0.1048 #####
  0.6000000000000001  0.0895 ####
  0.7000000000000001  0.0713 ###
  0.8                 0.0503 ##
  0.9                 0.0265 #
  1.0                 0.0

  scala> bpdf.toPmf(0.2).printChart()
  -1.0                 0.0
  -0.8                 0.0545 ##
  -0.6                 0.0969 ####
  -0.3999999999999999  0.1272 ######
  -0.19999999999999996 0.1454 #######
  0.0                  0.1515 #######
  0.20000000000000018  0.1454 #######
  0.40000000000000013  0.1272 ######
  0.6000000000000001   0.0969 ####
  0.8                  0.0545 ##
  1.0                  0.0

The Distributions extension provides methods for creating common Pdf such as Gaussian or Exponential distributions.

Extensions

This library was designed such that only the core operations needed for the creation and manipulation of the structures presented above are included in the class themselves. Additional features can be added by importing modules from the package extensions.

Plotting

The Plotting module provides support for graphical plotting, leveraging the powerful JFreeChart library with a custom theme. Pmf, Suite, Cdf and BoundedPdf instances can be plotted, as long as their keys have an Ordering (for plotting bar charts) or Numeric (for plotting XY line charts) implicit in scope:

  scala> import thinkbayes.extensions.Plotting._
  import thinkbayes.extensions.Plotting._

  scala> val xyChart = bpdf.plotXY("-x^2 + 1")
  xyChart: scalax.chart.XYChart = scalax.chart.ChartFactories$XYLineChart$$anon$17@290e640d

plotxy

  scala> val barChart = prior.plotBar("prior")
  barChart: scalax.chart.CategoryChart = scalax.chart.ChartFactories$BarChart$$anon$3@5c3e1ebe

plotbar_prior

New series can be added to a previously created chart. This is useful for comparing differences between two distributions or Bayesian suites:

  scala> posterior.plotBarOn(barChart, "after a 6 is rolled")
  res17: barChart.type = scalax.chart.ChartFactories$BarChart$$anon$3@5c3e1ebe

plotbar_posterior

Other attributes of the chart, such as the title and the axis labels, can be optionally specified.

Distributions

The Distributions module provides integration with the distribution implementations from Apache Commons Math, as well as several methods for creating Pmf and Pdf instances for common distributions:

  scala> import thinkbayes.extensions.Distributions._
  import thinkbayes.extensions.Distributions._

  scala> poissonPmf(3.0).plotBar("")
  res18: scalax.chart.CategoryChart = scalax.chart.ChartFactories$BarChart$$anon$3@6736cd9d

poisson

  scala> val tri: Pdf = new org.apache.commons.math3.distribution.TriangularDistribution(0.0, 0.5, 2.0)
  tri: thinkbayes.Pdf = thinkbayes.extensions.Distributions$$anon$1@7b5cdeb6

  scala> tri.bounded(0.0, 2.0).plotXY("")
  res19: scalax.chart.XYChart = scalax.chart.ChartFactories$XYLineChart$$anon$17@55a7c8a3

triangular

Finally, we can estimate a Pdf from a sequence of samples using kernel density estimation:

  scala> estimatePdf(Seq(1, 2, 2, 4, 4, 4, 9, 9, 9, 9, 11, 11, 15, 19)).bounded(0, 20).plotXY("")
  res20: scalax.chart.XYChart = scalax.chart.ChartFactories$XYLineChart$$anon$17@1c15725

kde

Stats

The Stats module is a simple extension that provides the calculation of quantiles and credible intervals to Pmf and Cdf instances:

  scala> import thinkbayes.extensions.Stats._
  import thinkbayes.extensions.Stats._

  scala> normalPmf(2.5, 1.5).quantile(0.5)
  res21: Double = 2.5

  scala> normalPmf(0.0, 1.0).credibleInterval(0.9)
  res22: (Double, Double) = (-1.6440000000000001,1.6440000000000001)

Sampling

Using Pmf merging methods such as mixture or join yield results as accurate as they can be, but they are also computationally expensive. The Sampling module aims to provide probabilistic alternatives based on sampling, which can be the only choice for large Pmf:

  scala> val dieList = Seq.fill(100)(die(6)) // a hundred dice
  dieList: Seq[thinkbayes.Pmf[Int]] = List(Map(5 -> 0.16666666666666666, 1 -> 0.16666666666666666, 6 -> 0.16666666666666666, 2 -> 0.16666666666666666, 3 -> 0.1666666666666666, 4 -> 0.16666666666666666),...

  scala> val xyChart = dieList.reduce(_ ++ _).plotXY("exact")
  xyChart: scalax.chart.XYChart = scalax.chart.ChartFactories$XYLineChart$$anon$17@30015846

  scala> sampleSum(dieList, 10000).plotXYOn(xyChart, "sampled")
  res23: xyChart.type = scalax.chart.ChartFactories$XYLineChart$$anon$17@81f0a53

sampling

Examples

A number of examples and problems explored throughout Think Bayes are implemented in the package examples in the test directory. They are always accompanied by the original problem description and I made an effort to make the steps of each problem as clear as possible.

Copyright

Copyright (c) 2014-2017 Rui Gonçalves. See LICENSE for details.

About

A Scala library for Bayesian Inference and Probabilistic Programming

Resources

License

Stars

Watchers

Forks

Languages