A probability monad for the bootstrap particle filter

Introduction

In the previous post I showed how to write your own general-purpose monadic probabilistic programming language from scratch in 50 lines of (Scala) code. That post is a pre-requisite for this one, so if you haven’t read it, go back and have a quick skim through it before proceeding. In that post I tried to keep everything as simple as possible, but at the expense of both elegance and efficiency. In this post I’ll address one problem with the implementation from that post – the memory (and computational) overhead associated with forming the Cartesian product of particle sets during monadic binding (flatMap). So if particle sets are typically of size , then the Cartesian product is of size , and multinomial resampling is applied to this set of size in order to sample back down to a set of size . But this isn’t actually necessary. We can directly construct a set of size , certainly saving memory, but also potentially saving computation time if the conditional distribution (on the right of the monadic bind) can be efficiently sampled. If we do this we will have a probability monad encapsulating the logic of a bootstrap particle filter, such as is often used for computing the filtering distribution of a state-space model in time series analysis. This simple change won’t solve the computational issues associated with deep monadic binding, but does solve the memory problem, and can lead to computationally efficient algorithms so long as care is taken in the formulation of probabilistic programs to ensure that deep monadic binding doesn’t occur. We’ll discuss that issue in the context of state-space models later, once we have our new SMC-based probability monad.

Materials for this post can be found in my blog repo, and a draft of this post itself can be found in the form of an executable tut document.

An SMC-based monad

The idea behind the approach to binding used in this monad is to mimic the “predict” step of a bootstrap particle filter. Here, for each particle in the source distribution, exactly one particle is drawn from the required conditional distribution and paired with the source particle, preserving the source particle’s original weight. So, in order to operationalise this, we will need a draw method adding into our probability monad. It will also simplify things to add a flatMap method to our Particle type constructor.

To follow along, you can type sbt console from the min-ppl2 directory of my blog repo, then paste blocks of code one at a time.

  import breeze.stats.{distributions => bdist}
  import breeze.linalg.DenseVector
  import cats._
  import cats.implicits._

  implicit val numParticles = 2000

  case class Particle[T](v: T, lw: Double) { // value and log-weight
    def map[S](f: T => S): Particle[S] = Particle(f(v), lw)
    def flatMap[S](f: T => Particle[S]): Particle[S] = {
      val ps = f(v)
      Particle(ps.v, lw + ps.lw)
    }
  }

I’ve added a dependence on cats here, so that we can use some derived methods, later. To take advantage of this, we must provide evidence that our custom types conform to standard type class interfaces. For example, we can provide evidence that Particle[_] is a monad as follows.

  implicit val particleMonad = new Monad[Particle] {
    def pure[T](t: T): Particle[T] = Particle(t, 0.0)
    def flatMap[T,S](pt: Particle[T])(f: T => Particle[S]): Particle[S] = pt.flatMap(f)
    def tailRecM[T,S](t: T)(f: T => Particle[Either[T,S]]): Particle[S] = ???
  }

The technical details are not important for this post, but we’ll see later what this can give us.

We can now define our Prob[_] monad in the following way.

  trait Prob[T] {
    val particles: Vector[Particle[T]]
    def draw: Particle[T]
    def mapP[S](f: T => Particle[S]): Prob[S] = Empirical(particles map (_ flatMap f))
    def map[S](f: T => S): Prob[S] = mapP(v => Particle(f(v), 0.0))
    def flatMap[S](f: T => Prob[S]): Prob[S] = mapP(f(_).draw)
    def resample(implicit N: Int): Prob[T] = {
      val lw = particles map (_.lw)
      val mx = lw reduce (math.max(_,_))
      val rw = lw map (lwi => math.exp(lwi - mx))
      val law = mx + math.log(rw.sum/(rw.length))
      val ind = bdist.Multinomial(DenseVector(rw.toArray)).sample(N)
      val newParticles = ind map (i => particles(i))
      Empirical(newParticles.toVector map (pi => Particle(pi.v, law)))
    }
    def cond(ll: T => Double): Prob[T] = mapP(v => Particle(v, ll(v)))
    def empirical: Vector[T] = resample.particles.map(_.v)
  }

  case class Empirical[T](particles: Vector[Particle[T]]) extends Prob[T] {
    def draw: Particle[T] = {
      val lw = particles map (_.lw)
      val mx = lw reduce (math.max(_,_))
      val rw = lw map (lwi => math.exp(lwi - mx))
      val law = mx + math.log(rw.sum/(rw.length))
      val idx = bdist.Multinomial(DenseVector(rw.toArray)).draw
      Particle(particles(idx).v, law)
    }
  }

As before, if you are pasting code blocks into the REPL, you will need to use :paste mode to paste these two definitions together.

The essential structure is similar to that from the previous post, but with a few notable differences. Most fundamentally, we now require any concrete implementation to provide a draw method returning a single particle from the distribution. Like before, we are not worrying about purity of functional code here, and using a standard random number generator with a globally mutating state. We can define a mapP method (for “map particle”) using the new flatMap method on Particle, and then use that to define map and flatMap for Prob[_]. Crucially, draw is used to define flatMap without requiring a Cartesian product of distributions to be formed.

We add a draw method to our Empirical[_] implementation. This method is computationally intensive, so it will still be computationally problematic to chain several flatMaps together, but this will no longer be in memory utilisation. Note that again we carefully set the weight of the drawn particle so that its raw weight is the average of the raw weight of the empirical distribution. This is needed to propagate conditioning information correctly back through flatMaps. There is obviously some code duplication between the draw method on Empirical and the resample method on Prob, but I’m not sure it’s worth factoring out.

It is worth noting that neither flatMap nor cond triggers resampling, so the user of the library is now responsible for resampling when appropriate.

We can provide evidence that Prob[_] forms a monad just like we did Particle[_].

  implicit val probMonad = new Monad[Prob] {
    def pure[T](t: T): Prob[T] = Empirical(Vector(Particle(t, 0.0)))
    def flatMap[T,S](pt: Prob[T])(f: T => Prob[S]): Prob[S] = pt.flatMap(f)
    def tailRecM[T,S](t: T)(f: T => Prob[Either[T,S]]): Prob[S] = ???
  }

Again, we’ll want to be able to create a distribution from an unweighted collection of values.

  def unweighted[T](ts: Vector[T], lw: Double = 0.0): Prob[T] =
    Empirical(ts map (Particle(_, lw)))

We will again define an implementation for distributions with tractable likelihoods, which are therefore easy to condition on. They will typically also be easy to draw from efficiently, and we will use this fact, too.

  trait Dist[T] extends Prob[T] {
    def ll(obs: T): Double
    def ll(obs: Seq[T]): Double = obs map (ll) reduce (_+_)
    def fit(obs: Seq[T]): Prob[T] = mapP(v => Particle(v, ll(obs)))
    def fitQ(obs: Seq[T]): Prob[T] = Empirical(Vector(Particle(obs.head, ll(obs))))
    def fit(obs: T): Prob[T] = fit(List(obs))
    def fitQ(obs: T): Prob[T] = fitQ(List(obs))
  }

We can give implementations of this for a few standard distributions.

  case class Normal(mu: Double, v: Double)(implicit N: Int) extends Dist[Double] {
    lazy val particles = unweighted(bdist.Gaussian(mu, math.sqrt(v)).
      sample(N).toVector).particles
    def draw = Particle(bdist.Gaussian(mu, math.sqrt(v)).draw, 0.0)
    def ll(obs: Double) = bdist.Gaussian(mu, math.sqrt(v)).logPdf(obs)
  }

  case class Gamma(a: Double, b: Double)(implicit N: Int) extends Dist[Double] {
    lazy val particles = unweighted(bdist.Gamma(a, 1.0/b).
      sample(N).toVector).particles
    def draw = Particle(bdist.Gamma(a, 1.0/b).draw, 0.0)
    def ll(obs: Double) = bdist.Gamma(a, 1.0/b).logPdf(obs)
  }

  case class Poisson(mu: Double)(implicit N: Int) extends Dist[Int] {
    lazy val particles = unweighted(bdist.Poisson(mu).
      sample(N).toVector).particles
    def draw = Particle(bdist.Poisson(mu).draw, 0.0)
    def ll(obs: Int) = bdist.Poisson(mu).logProbabilityOf(obs)
  }

Note that we now have to provide an (efficient) draw method for each implementation, returning a single draw from the distribution as a Particle with a (raw) weight of 1 (log weight of 0).

We are done. It’s a few more lines of code than that from the previous post, but this is now much closer to something that could be used in practice to solve actual inference problems using a reasonable number of particles. But to do so we will need to be careful do avoid deep monadic binding. This is easiest to explain with a concrete example.

Using the SMC-based probability monad in practice

Monadic binding and applicative structure

As explained in the previous post, using Scala’s for-expressions for monadic binding gives a natural and elegant PPL for our probability monad “for free”. This is fine, and in general there is no reason why using it should lead to inefficient code. However, for this particular probability monad implementation, it turns out that deep monadic binding comes with a huge performance penalty. For a concrete example, consider the following specification, perhaps of a prior distribution over some independent parameters.

    val prior = for {
      x <- Normal(0,1)
      y <- Gamma(1,1)
      z <- Poisson(10)
    } yield (x,y,z)

Don’t paste that into the REPL – it will take an age to complete!

Again, I must emphasise that there is nothing wrong with this specification, and there is no reason in principle why such a specification can’t be computationally efficient – it’s just a problem for our particular probability monad. We can begin to understand the problem by thinking about how this will be de-sugared by the compiler. Roughly speaking, the above will de-sugar to the following nested flatMaps.

    val prior2 =
      Normal(0,1) flatMap {x =>
        Gamma(1,1) flatMap {y =>
          Poisson(10) map {z =>
            (x,y,z)}}}

Again, beware of pasting this into the REPL.

So, although written from top to bottom, the nesting is such that the flatMaps collapse from the bottom-up. The second flatMap (the first to collapse) isn’t such a problem here, as the Poisson has a draw method. But the result is an empirical distribution, which has an draw method. So the first flatMap (the second to collapse) is an operation. By extension, it’s easy to see that the computational cost of nested flatMaps will be exponential in the number of monadic binds. So, looking back at the for expression, the problem is that there are three <-. The last one isn’t a problem since it corresponds to a map, and the second last one isn’t a problem, since the final distribution is tractable with an draw method. The problem is the first <-, corresponding to a flatMap of one empirical distribution with respect to another. For our particular probability monad, it’s best to avoid these if possible.

The interesting thing to note here is that because the distributions are independent, there is no need for them to be sequenced. They could be defined in any order. In this case it makes sense to use the applicative structure implied by the monad.

Now, since we have told cats that Prob[_] is a monad, it can provide appropriate applicative methods for us automatically. In Cats, every monad is assumed to be also an applicative functor (which is true in Cartesian closed categories, and Cats implicitly assumes that all functors and monads are defined over CCCs). So we can give an alternative specification of the above prior using applicative composition.

 val prior3 = Applicative[Prob].tuple3(Normal(0,1), Gamma(1,1), Poisson(10))
// prior3: Wrapped.Prob[(Double, Double, Int)] = Empirical(Vector(Particle((-0.057088546468105204,0.03027578552505779,9),0.0), Particle((-0.43686658266043743,0.632210127012762,14),0.0), Particle((-0.8805715148936012,3.4799656228544706,4),0.0), Particle((-0.4371726407147289,0.0010707859994652403,12),0.0), Particle((2.0283297088320755,1.040984491158822,10),0.0), Particle((1.2971862986495886,0.189166705596747,14),0.0), Particle((-1.3111333817551083,0.01962422606642761,9),0.0), Particle((1.6573851896142737,2.4021836368401415,9),0.0), Particle((-0.909927220984726,0.019595551644771683,11),0.0), Particle((0.33888133893822464,0.2659823344145805,10),0.0), Particle((-0.3300797295729375,3.2714740256437667,10),0.0), Particle((-1.8520554352884224,0.6175322756460341,10),0.0), Particle((0.541156780497547...

This one is mathematically equivalent, but safe to paste into your REPL, as it does not involve deep monadic binding, and can be used whenever we want to compose together independent components of a probabilistic program. Note that “tupling” is not the only possibility – Cats provides a range of functions for manipulating applicative values.

This is one way to avoid deep monadic binding, but another strategy is to just break up a large for expression into separate smaller for expressions. We can examine this strategy in the context of state-space modelling.

Particle filtering for a non-linear state-space model

We can now re-visit the DGLM example from the previous post. We began by declaring some observations and a prior.

    val data = List(2,1,0,2,3,4,5,4,3,2,1)
// data: List[Int] = List(2, 1, 0, 2, 3, 4, 5, 4, 3, 2, 1)

    val prior = for {
      w <- Gamma(1, 1)
      state0 <- Normal(0.0, 2.0)
    } yield (w, List(state0))
// prior: Wrapped.Prob[(Double, List[Double])] = Empirical(Vector(Particle((4.220683377724395,List(0.37256749723762683)),0.0), Particle((0.4436668049925418,List(-1.0053578391265572)),0.0), Particle((0.9868899648436931,List(-0.6985099310193449)),0.0), Particle((0.13474375773634908,List(0.9099291736792412)),0.0), Particle((1.9654021747685184,List(-0.042127103727998175)),0.0), Particle((0.21761202474220223,List(1.1074616830012525)),0.0), Particle((0.31037163527711015,List(0.9261849914020324)),0.0), Particle((1.672438830781466,List(0.01678529855289384)),0.0), Particle((0.2257151759143097,List(2.5511304854128354)),0.0), Particle((0.3046489890769499,List(3.2918304533361398)),0.0), Particle((1.5115941814057159,List(-1.633612165168878)),0.0), Particle((1.4185906813831506,List(-0.8460922678989864))...

Looking carefully at the for-expression, there are just two <-, and the distribution on the RHS of the flatMap is tractable, so this is just . So far so good.

Next, let’s look at the function to add a time point, which previously looked something like the following.

    def addTimePointSIS(current: Prob[(Double, List[Double])],
      obs: Int): Prob[(Double, List[Double])] = {
      println(s"Conditioning on observation: $obs")
      for {
        tup <- current
        (w, states) = tup
        os = states.head
        ns <- Normal(os, w)
        _ <- Poisson(math.exp(ns)).fitQ(obs)
      } yield (w, ns :: states)
    }
// addTimePointSIS: (current: Wrapped.Prob[(Double, List[Double])], obs: Int)Wrapped.Prob[(Double, List[Double])]

Recall that our new probability monad does not automatically trigger resampling, so applying this function in a fold will lead to a simple sampling importance sampling (SIS) particle filter. Typically, the bootstrap particle filter includes resampling after each time point, giving a special case of a sampling importance resampling (SIR) particle filter, which we could instead write as follows.

    def addTimePointSimple(current: Prob[(Double, List[Double])],
      obs: Int): Prob[(Double, List[Double])] = {
      println(s"Conditioning on observation: $obs")
      val updated = for {
        tup <- current
        (w, states) = tup
        os = states.head
        ns <- Normal(os, w)
        _ <- Poisson(math.exp(ns)).fitQ(obs)
      } yield (w, ns :: states)
      updated.resample
    }
// addTimePointSimple: (current: Wrapped.Prob[(Double, List[Double])], obs: Int)Wrapped.Prob[(Double, List[Double])]

This works fine, but we can see that there are three <- in this for expression. This leads to a flatMap with an empirical distribution on the RHS, and hence is . But this is simple enough to fix, by separating the updating process into separate “predict” and “update” steps, which is how people typically formulate particle filters for state-space models, anyway. Here we could write that as

    def addTimePoint(current: Prob[(Double, List[Double])],
      obs: Int): Prob[(Double, List[Double])] = {
      println(s"Conditioning on observation: $obs")
      val predict = for {
        tup <- current
        (w, states) = tup
        os = states.head
        ns <- Normal(os, w)
      }
      yield (w, ns :: states)
      val updated = for {
        tup <- predict
        (w, states) = tup
        st = states.head
        _ <- Poisson(math.exp(st)).fitQ(obs)
      } yield (w, states)
      updated.resample
    }
// addTimePoint: (current: Wrapped.Prob[(Double, List[Double])], obs: Int)Wrapped.Prob[(Double, List[Double])]

By breaking the for expression into two: the first for the “predict” step and the second for the “update” (conditioning on the observation), we get two operations, which for large is clearly much faster. We can then run the filter by folding over the observations.

  import breeze.stats.{meanAndVariance => meanVar}
// import breeze.stats.{meanAndVariance=>meanVar}

  val mod = data.foldLeft(prior)(addTimePoint(_,_)).empirical
// Conditioning on observation: 2
// Conditioning on observation: 1
// Conditioning on observation: 0
// Conditioning on observation: 2
// Conditioning on observation: 3
// Conditioning on observation: 4
// Conditioning on observation: 5
// Conditioning on observation: 4
// Conditioning on observation: 3
// Conditioning on observation: 2
// Conditioning on observation: 1
// mod: Vector[(Double, List[Double])] = Vector((0.24822528144246606,List(0.06290285371838457, 0.01633338109272575, 0.8997103339551227, 1.5058726341571411, 1.0579925693609091, 1.1616536515200064, 0.48325623593870665, 0.8457351097543767, -0.1988290999293708, -0.4787511341321954, -0.23212497417019512, -0.15327432440577277)), (1.111430233331792,List(0.6709342824443849, 0.009092797044165657, -0.13203367846117453, 0.4599952735399485, 1.3779288637042504, 0.6176597963402879, 0.6680455419800753, 0.48289163013446945, -0.5994001698510807, 0.4860969602653898, 0.10291798193078927, 1.2878325765987266)), (0.6118925941009055,List(0.6421161986636132, 0.679470360928868, 1.0552459559203342, 1.200835166087372, 1.3690372269589233, 1.8036766847282912, 0.6229883551656629, 0.14872642198313774, -0.122700856878725...

  meanVar(mod map (_._1)) // w
// res0: breeze.stats.meanAndVariance.MeanAndVariance = MeanAndVariance(0.2839184023932576,0.07391602428256917,2000)

  meanVar(mod map (_._2.reverse.head)) // initial state
// res1: breeze.stats.meanAndVariance.MeanAndVariance = MeanAndVariance(0.26057368528422714,0.4802810202354611,2000)

  meanVar(mod map (_._2.head)) // final state
// res2: breeze.stats.meanAndVariance.MeanAndVariance = MeanAndVariance(0.5448036669181697,0.28293080584600894,2000)

Summary and conclusions

Here we have just done some minor tidying up of the rather naive probability monad from the previous post to produce an SMC-based probability monad with improved performance characteristics. Again, we get an embedded probabilistic programming language “for free”. Although the language itself is very flexible, allowing us to construct more-or-less arbitrary probabilistic programs for Bayesian inference problems, we saw that a bug/feature of this particular inference algorithm is that care must be taken to avoid deep monadic binding if reasonable performance is to be obtained. In most cases this is simple to achieve by using applicative composition or by breaking up large for expressions.

There are still many issues and inefficiencies associated with this PPL. In particular, if the main intended application is to state-space models, it would make more sense to tailor the algorithms and implementations to exactly that case. OTOH, if the main concern is a generic PPL, then it would make sense to make the PPL independent of the particular inference algorithm. These are both potential topics for future posts.

Software

• min-ppl2 – code associated with this blog post
• Rainier – a more efficient PPL with similar syntax
• monad-bayes – a Haskell library exploring related ideas

A scalable particle filter in Scala

Introduction

Many modern algorithms in computational Bayesian statistics have at their heart a particle filter or some other sequential Monte Carlo (SMC) procedure. In this blog I’ve discussed particle MCMC algorithms which use a particle filter in the inner-loop in order to compute a (noisy, unbiased) estimate of the marginal likelihood of the data. These algorithms are often very computationally intensive, either because the forward model used to propagate the particles is expensive, or because the likelihood associated with each particle/observation is expensive (or both). In this case it is desirable to parallelise the particle filter to run on all available cores of a machine, or in some cases, it would even be desirable to distribute the the particle filter computation across a cluster of machines.

Parallelisation is difficult when using the conventional imperative programming languages typically used in scientific and statistical computing, but is much easier using modern functional languages such as Scala. In fact, in languages such as Scala it is possible to describe algorithms at a higher level of abstraction, so that exactly the same algorithm can run in serial, run in parallel across all available cores on a single machine, or run in parallel across a cluster of machines, all without changing any code. Doing so renders parallelisation a non-issue. In this post I’ll talk through how to do this for a simple bootstrap particle filter, but the same principle applies for a large range of statistical computing algorithms.

Typeclasses and monadic collections

In the previous post I gave a quick introduction to the monad concept, and to monadic collections in particular. Many computational tasks in statistics can be accomplished using a sequence of operations on monadic collections. We would like to write code that is independent of any particular implementation of a monadic collection, so that we can switch to a different implementation without changing the code of our algorithm (for example, switching from a serial to a parallel collection). But in strongly typed languages we need to know at compile time that the collection we use has the methods that we require. Typeclasses provide a nice solution to this problem. I don’t want to get bogged down in a big discussion about Scala typeclasses here, but suffice to say that they describe a family of types conforming to a particular interface in an ad hoc loosely coupled way (they are said to provide ad hoc polymorphism). They are not the same as classes in traditional O-O languages, but they do solve a similar problem to the adaptor design pattern, in a much cleaner way. We can describe a simple typeclass for our monadic collection as follows:

trait GenericColl[C[_]] {
  def map[A, B](ca: C[A])(f: A => B): C[B]
  def reduce[A](ca: C[A])(f: (A, A) => A): A
  def flatMap[A, B, D[B] <: GenTraversable[B]](ca: C[A])(f: A => D[B]): C[B]
  def zip[A, B](ca: C[A])(cb: C[B]): C[(A, B)]
  def length[A](ca: C[A]): Int
}

In the typeclass we just list the methods that we expect our generic collection to provide, but do not say anything about how they are implemented. For example, we know that operations such as map and reduce can be executed in parallel, but this is a separate concern. We can now write code that can be used for any collection conforming to the requirements of this typeclass. The full code for this example is provided in the associated github repo for this blog, and includes the obvious syntax for this typeclass, and typeclass instances for the Scala collections Vector and ParVector, that we will exploit later in the example.

SIR step for a bootstrap filter

We can now write some code for a single observation update of a bootstrap particle filter.

def update[S: State, O: Observation, C[_]: GenericColl](
  dataLik: (S, O) => LogLik, stepFun: S => S
)(x: C[S], o: O): (LogLik, C[S]) = {
  val xp = x map (stepFun(_))
  val lw = xp map (dataLik(_, o))
  val max = lw reduce (math.max(_, _))
  val rw = lw map (lwi => math.exp(lwi - max))
  val srw = rw reduce (_ + _)
  val l = rw.length
  val z = rw zip xp
  val rx = z flatMap (p => Vector.fill(Poisson(p._1 * l / srw).draw)(p._2))
  (max + math.log(srw / l), rx)
}

This is a very simple bootstrap filter, using Poisson resampling for simplicity and data locality, but does include use of the log-sum-exp trick to prevent over/underflow of raw weight calculations, and tracks the marginal (log-)likelihood of the observation. With this function we can now pass in a “prior” particle distribution in any collection conforming to our typeclass, together with a propagator function, an observation (log-)likelihood, and an observation, and it will return back a new collection of particles of exactly the same type that was provided for input. Note that all of the operations we require can be accomplished with the standard monadic collection operations declared in our typeclass.

Filtering as a functional fold

Once we have a function for executing one step of a particle filter, we can produce a function for particle filtering as a functional fold over a sequence of observations:

def pFilter[S: State, O: Observation, C[_]: GenericColl, D[O] <: GenTraversable[O]](
  x0: C[S], data: D[O], dataLik: (S, O) => LogLik, stepFun: S => S
): (LogLik, C[S]) = {
  val updater = update[S, O, C](dataLik, stepFun) _
  data.foldLeft((0.0, x0))((prev, o) => {
    val next = updater(prev._2, o)
    (prev._1 + next._1, next._2)
  })
}

Folding data structures is a fundamental concept in functional programming, and is exactly what is required for any kind of filtering problem. Note that Brian Beckman has recently written a series of articles on Kalman filtering as a functional fold.

Marginal likelihoods and parameter estimation

So far we haven’t said anything about parameters or parameter estimation, but this is appropriate, since parametrisation is a separate concern from filtering. However, once we have a function for particle filtering, we can produce a function concerned with evaluating marginal likelihoods trivially:

def pfMll[S: State, P: Parameter, O: Observation, 
            C[_]: GenericColl, D[O] <: GenTraversable[O]](
  simX0: P => C[S], stepFun: P => S => S, 
  dataLik: P => (S, O) => LogLik, data: D[O]
): (P => LogLik) = (th: P) => 
       pFilter(simX0(th), data, dataLik(th), stepFun(th))._1

Note that this higher-order function does not return a value, but instead a function which will accept a parameter as input and return a (log-)likelihood as output. This can then be used for parameter estimation purposes, perhaps being used in a PMMH pMCMC algorithm, or something else. Again, this is a separate concern.

Example

Here I’ll just give a completely trivial toy example, purely to show how the functions work. For avoidance of doubt, I know that there are many better/simpler/easier ways to tackle this problem! Here we will just look at inferring the auto-regression parameter of a linear Gaussian AR(1)-plus-noise model using the functions we have developed.

First we can simulate some synthetic data from this model, using a value of 0.8 for the auto-regression parameter:

val inNoise = Gaussian(0.0, 1.0).sample(99)
val state = DenseVector(inNoise.scanLeft(0.0)((s, i) => 0.8 * s + i).toArray)
val noise = DenseVector(Gaussian(0.0, 2.0).sample(100).toArray)
val data = (state + noise).toArray.toList

Now assuming that we don’t know the auto-regression parameter, we can construct a function to evaluate the likelihood of different parameter values as follows:

val mll = pfMll(
  (th: Double) => Gaussian(0.0, 10.0).sample(10000).toVector.par,
  (th: Double) => (s: Double) => Gaussian(th * s, 1.0).draw,
  (th: Double) => (s: Double, o: Double) => Gaussian(s, 2.0).logPdf(o),
  data
)

Note that the 4 characters “.par” at the end of line 2 are the only difference between running this code serially or in parallel! Now we can run this code by calling the returned function with different values. So, hopefully mll(0.8) will return a larger log-likelihood than (say) mll(0.6) or mll(0.9). The example code in the github repo plots the results of calling mll() for a range of values (note that if that was the genuine use-case, then it would be much better to parallellise the parameter range than the particle filter, due to providing better parallelisation granularity, but many other examples require parallelisation of the particle filter itself). In this particular example, both the forward model and the likelihood are very cheap operations, so there is little to be gained from parallelisation. Nevertheless, I still get a speedup of more than a factor of two using the parallel version on my laptop.

Conclusion

In this post we have shown how typeclasses can be used in Scala to write code that is parallelisation-agnostic. Code written in this way can be run on one or many cores as desired. We’ve illustrated the concept with a scalable particle filter, but nothing about the approach is specific to that application. It would be easy to build up a library of statistical routines this way, all of which can effectively exploit available parallel hardware. Further, although we haven’t demonstrated it here, it is trivial to extend this idea to allow code to be distribution over a cluster of parallel machines if necessary. For example, if an Apache Spark cluster is available, it is easy to make a Spark RDD instance for our generic collection typeclass, that will then allow us to run our (unmodified) particle filter code over a Spark cluster. This emphasises the fact that Spark can be useful for distributing computation as well as just processing “big data”. I’ll say more about Spark in subsequent posts.

Tuning particle MCMC algorithms

Several papers have appeared recently discussing the issue of how to tune the number of particles used in the particle filter within a particle MCMC algorithm such as particle marginal Metropolis Hastings (PMMH). Three such papers are:

• Doucet, Arnaud, Michael Pitt, and Robert Kohn. Efficient implementation of Markov chain Monte Carlo when using an unbiased likelihood estimator. arXiv preprint arXiv:1210.1871 (2012).
• Pitt, Michael K., Ralph dos Santos Silva, Paolo Giordani, and Robert Kohn. On some properties of Markov chain Monte Carlo simulation methods based on the particle filter. Journal of Econometrics 171, no. 2 (2012): 134-151.
• Sherlock, Chris, Alexandre H. Thiery, Gareth O. Roberts, and Jeffrey S. Rosenthal. On the efficiency of pseudo-marginal random walk Metropolis algorithms. arXiv preprint arXiv:1309.7209 (2013).

I have discussed psuedo marginal MCMC and particle MCMC algorithms in previous posts. It will be useful to refer back to these posts if these topics are unfamiliar. Within particle MCMC algorithms (and psuedo-marginal MCMC algorithms, more generally), an unbiased estimate of marginal likelihood is constructed using a number of particles. The more particles that are used, the better the estimate of marginal likelihood is, and the resulting MCMC algorithm will behave more like a “real” marginal MCMC algorithm. For a small number of particles, the algorithm will still have exactly the correct target, but the noise in the unbiased estimator of marginal likelihood will lead to poor mixing of the MCMC chain. The idea is to use just enough particles to ensure that there isn’t “too much” noise in the unbiased estimator, but not to waste lots of time producing a super-accurate estimate of marginal likelihood if that isn’t necessary to ensure good mixing of the MCMC chain.

The papers above try to give theoretical justifications for certain “rules of thumb” that are commonly used in practice. One widely adopted scheme is to tune the number of particles so that the variance of the log of the estimate of marginal liklihood is around one. The obvious questions are “where?” and “why?”, and these questions turn out to be connected. As we will see, there isn’t really a good answer to the “where?” question, but what people usually do is use a pilot run to get an estimate of the posterior mean, or mode, or MLE, and then pick one and tune the noise variance at that particular parameter value. As to “why?”, well, the papers above make various (slightly different) assumptions, all of which lead to trading off mixing against computation time to obtain an “optimal” number of particles. They don’t all agree that the variance of the noise should be exactly 1, but they all agree to an order of magnitude.

All of the above papers make the assumption that the noise distribution associated with the marginal likelihood estimate is independent of the parameter at which it is being evaluated, which explains why there isn’t a really good answer to the “where?” question – under the assumption it doesn’t matter what parameter value is used for tuning – they are all the same! Easy. Except that’s quite a big assumption, so it would be nice to know that it is reasonable, and unfortunately it isn’t. Let’s look at an example to see what goes wrong.

Example

In Chapter 10 of my book I look in detail at constructing a PMMH algorithm for inferring the parameters of a discretely observed stochastic Lotka-Volterra model. I’ve stepped through the computational details in a previous post which you should refer back to for the necessary background. Following that post, we can construct a particle filter to return an unbiased estimate of marginal likelihood using the following R code (which relies on the smfsb CRAN package):

require(smfsb)
# data
data(LVdata)
data=as.timedData(LVnoise10)
noiseSD=10
# measurement error model
dataLik <- function(x,t,y,log=TRUE,...)
{
    ll=sum(dnorm(y,x,noiseSD,log=TRUE))
    if (log)
        return(ll)
    else
        return(exp(ll))
}
# now define a sampler for the prior on the initial state
simx0 <- function(N,t0,...)
{
    mat=cbind(rpois(N,50),rpois(N,100))
    colnames(mat)=c("x1","x2")
    mat
}
# construct particle filter
mLLik=pfMLLik(150,simx0,0,stepLVc,dataLik,data)

Again, see the relevant previous post for details. So now mLLik() is a function that will return the log of an unbiased estimate of marginal likelihood (based on 150 particles) given a parameter value at which to evaluate.

What we are currently wondering is whether the noise in the estimate is independent of the parameter at which it is evaluated. We can investigate this for this filter easily by looking at how the estimate varies as the first parameter (prey birth rate) varies. The following code computes a log likelihood estimate across a range of values and plots the result.

mLLik1=function(x){mLLik(th=c(th1=x,th2=0.005,th3=0.6))}
x=seq(0.7,1.3,length.out=5001)
y=sapply(x,mLLik1)
plot(x[y>-1e10],y[y>-1e10])

The resulting plot is as follows (click for full size):

So, looking at the plot, it is very clear that the noise variance certainly isn’t constant as the parameter varies – it varies substantially. Furthermore, the way in which it varies is “dangerous”, in that the noise is smallest in the vicinity of the MLE. So, if a parameter close to the MLE is chosen for tuning the number of particles, this will ensure that the noise is small close to the MLE, but not elsewhere in parameter space. This could have bad consequences for the mixing of the MCMC algorithm as it explores the tails of the posterior distribution.

So with the above in mind, how should one tune the number of particles in a pMCMC algorithm? I can’t give a general answer, but I can explain what I do. We can’t rely on theory, so a pragmatic approach is required. The above rule of thumb usually gives a good starting point for exploration. Then I just directly optimise ESS per CPU second of the pMCMC algorithm from pilot runs for varying numbers of particles (and other tuning parameters in the algorithm). ESS is “expected sample size”, which can be estimated using the effectiveSize() function in the coda CRAN package. Ugly and brutish, but it works…

Introduction to the particle Gibbs sampler

Introduction

Particle MCMC (the use of approximate SMC proposals within exact MCMC algorithms) is arguably one of the most important developments in computational Bayesian inference of the 21st Century. The key concepts underlying these methods are described in a famously impenetrable “read paper” by Andrieu et al (2010). Probably the most generally useful method outlined in that paper is the particle marginal Metropolis-Hastings (PMMH) algorithm that I have described previously – that post is required preparatory reading for this one.

In this post I want to discuss some of the other topics covered in the pMCMC paper, leading up to a description of the particle Gibbs sampler. The basic particle Gibbs algorithm is arguably less powerful than PMMH for a few reasons, some of which I will elaborate on. But there is still a lot of active research concerning particle Gibbs-type algorithms, which are attempting to address some of the deficiencies of the basic approach. Clearly, in order to understand and appreciate the recent developments it is first necessary to understand the basic principles, and so that is what I will concentrate on here. I’ll then finish with some pointers to more recent work in this area.

PIMH

I will adopt the same approach and notation as for my post on the PMMH algorithm, using a simple bootstrap particle filter for a state space model as the SMC proposal. It is simplest to understand particle Gibbs first in the context of known static parameters, and so it is helpful to first reconsider the special case of the PMMH algorithm where there are no unknown parameters and only the state path, of the process is being updated. That is, we target (for known, fixed, ) rather than . This special case is known as the particle independent Metropolis-Hastings (PIMH) sampler.

Here we envisage proposing a new path using a bootstrap filter, and then accepting the proposal with probability , where is the Metropolis-Hastings ratio

where is the bootstrap filter’s estimate of marginal likelihood for the new path, and is the estimate associated with the current path. Again using notation from the previous post it is clear that this ratio targets a distribution on the joint space of all simulated random variables proportional to

and that in this case the marginal distribution of the accepted path is exactly . Again, be sure to see the previous post for the explanation.

Conditional SMC update

So far we have just recapped the previous post in the case of known parameters, but it gives us insight in how to proceed. A general issue with Metropolis independence samplers in high dimensions is that they often exhibit “sticky” behaviour, whereby an unusually “good” accepted path is hard to displace. This motivates consideration of a block-Gibbs-style algorithm where updates are used that are always accepted. It is clear that simply running a bootstrap filter will target the particle filter distribution

and so the marginal distribution of the accepted path will be the approximate rather than the exact conditional distribution . However, we know from consideration of the PIMH algorithm that what we really want to do is target the slightly modified distribution proportional to

,

as this will lead to accepted paths with the exact marginal distribution. For the PIMH this modification is achieved using a Metropolis-Hastings correction, but we now try to avoid this by instead conditioning on the previously accepted path. For this target the accepted paths have exactly the required marginal distribution, so we now write the target as the product of the marginal for the current path times a conditional for all of the remaining variables.

where in addition to the correct marginal for we assume iid uniform ancestor indices. The important thing to note here is that the conditional distribution of the remaining variables simplifies to

.

The terms in the denominator are precisely the terms in the numerator corresponding to the current path, and hence “cancel out” the current path terms in the numerator. It is therefore clear that we can sample directly from this conditional distribution by running a bootstrap particle filter that includes the current path and which leaves the current path fixed. This is the conditional SMC (CSMC) update, which here is just a conditional bootstrap particle filter update. It is clear from the form of the conditional density how this filter must be constructed, but for completeness it is described below.

The bootstrap filter is run conditional on one trajectory. This is usually the trajectory sampled at the last run of the particle filter. The idea is that you do not sample new state or ancestor values for that one trajectory. Note that this guarantees that the conditioned on trajectory survives the filter right through to the final sweep of the filter at which point a new trajectory is picked from the current selection of paths, of which the conditioned-on trajectory is one.

Let be the path that is to be conditioned on, with ancestral lineage . Then, for , sample and set . Now suppose that at time we have a weighted sample from . First resample by sampling . Next sample . Then for all set and normalise with . Propagate this weighted set of particles to the next time point. At time select a single trajectory by sampling .

This defines a block Gibbs sampler which updates of the random variables in the augmented state space at each iteration. Since the block of variables to be updated is random, this defines an ergodic sampler for particles, and we have explained why the marginal distribution of the selected trajectory is the exact conditional distribution.

Before going on to consider the introduction of unknown parameters, it is worth considering the limitations of this method. One of the main motivations for considering a Gibbs-style update was concern about the “stickiness” of a Metropolis independence sampler. However, it is clear that conditional SMC updates also have the potential to stick. For a large number of time points, particle filter genealogies coalesce, or degenerate, to a single path. Since here we are conditioning on the current path, if there is coalescence, it is guaranteed to be to the previous path. So although the conditional SMC updates are always accepted, it is likely that much of the new path will be identical to the previous path, which is just another kind of “sticking” of the sampler. This problem with conditional SMC and particle Gibbs more generally is well recognised, and quite a bit of recent research activity in this area is directed at alleviating this sticking problem. The most obvious strategy to use is “backward sampling” (Godsill et al, 2004), which has been used in this context by Lindsten and Schon (2012), Whiteley et al (2010), and Chopin and Singh (2013), among others. Another related idea is “ancestor sampling” (Lindsten et al, 2014), which can be done in a single forward pass. Both of these techniques work well, but both rely on the tractability of the transition kernel of the state space model, which can be problematic in certain applications.

Particle Gibbs sampling

As we are working in the context of Gibbs-style updates, the introduction of static parameters, , into the problem is relatively straightforward. It turns out to be correct to do the obvious thing, which is to alternate between sampling given and the currently sampled path, , and sampling a new path using a conditional SMC update, conditional on the previous path in addition to and . Although this is the obvious thing to do, understanding exactly why it works is a little delicate, due to the augmented state space and conditional SMC update. However, it is reasonably clear that this strategy defines a “collapsed Gibbs sampler” (Lui, 1994), and so actually everything is fine. This particular collapsed Gibbs sampler is relatively easy to understand as a marginal sampler which integrates out the augmented variables, but then nevertheless samples the augmented variables at each iteration conditional on everything else.

Note that the Gibbs update of may be problematic in the context of a state space model with intractable transition kernel.

In a subsequent post I’ll show how to code up the particle Gibbs and other pMCMC algorithms in a reasonably efficient way.

References

• Andrieu, Christophe, Arnaud Doucet, and Roman Holenstein. “Particle Markov chain Monte Carlo methods.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 72.3 (2010): 269-342.
• Chopin, Nicolas, and Sumeetpal S. Singh. “On the particle Gibbs sampler.” arXiv preprint arXiv:1304.1887 (2013).
• Godsill, Simon J., Arnaud Doucet, and Mike West. “Monte Carlo smoothing for nonlinear time series.” Journal of the American Statistical Association 99.465 (2004).
• Lindsten, Fredrik, and Thomas B. Schön. “On the use of backward simulation in the particle Gibbs sampler.” Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on. IEEE, 2012.
• Lindsten, Fredrik, Michael I. Jordan, and Thomas B. Schön. “Particle Gibbs with Ancestor Sampling.” arXiv preprint arXiv:1401.0604 (2014).
• Liu, Jun S. “The collapsed Gibbs sampler in Bayesian computations with applications to a gene regulation problem.” Journal of the American Statistical Association 89.427 (1994): 958-966.
• Whiteley, Nick, Christophe Andrieu, and Arnaud Doucet. “Efficient Bayesian inference for switching state-space models using discrete particle Markov chain Monte Carlo methods.” arXiv preprint arXiv:1011.2437 (2010).

Parallel particle filtering and pMCMC using R and multicore

Introduction

In a previous post I showed how to construct a PMMH pMCMC algorithm for parameter estimation with partially observed Markov processes. The inner loop of a pMCMC algorithm consists of running a particle filter to construct an unbiased estimate of marginal likelihood. This inner loop is the place where the code spends almost all of its time, and so speeding up the particle filter will result in dramatic speedup of the pMCMC algorithm. This is fortunate, since as previously discussed, MCMC algorithms are difficult to parallelise other than on a per iteration basis. Here, each iteration can be speeded up if we can effectively parallelise a particle filter. Particle filters are much easier to parallelise than MCMC algorithms, and so it is tempting to try and exploit this within R. In fact, although it is the case that it is possible to effectively parallelise particle filters in efficient languages using low-level parallelisation tools (say, using C with MPI, or Java concurrency tools), it is not so easy to speed up R-based particle filters using R’s high-level parallelisation constructs, as we shall see.

Particle filters

In the previous post we looked at the function pfMLLik within the CRAN package smfsb. As a reminder, the source code is

pfMLLik <- function (n, simx0, t0, stepFun, dataLik, data) 
{
    times = c(t0, as.numeric(rownames(data)))
    deltas = diff(times)
    return(function(...) {
        xmat = simx0(n, t0, ...)
        ll = 0
        for (i in 1:length(deltas)) {
            xmat = t(apply(xmat, 1, stepFun, t0 = times[i], deltat = deltas[i], ...))
            w = apply(xmat, 1, dataLik, t = times[i + 1], y = data[i,], log = FALSE, ...)
            if (max(w) < 1e-20) {
                warning("Particle filter bombed")
                return(-1e+99)
            }
            ll = ll + log(mean(w))
            rows = sample(1:n, n, replace = TRUE, prob = w)
            xmat = xmat[rows, ]
        }
        ll
    })
}

The function itself doesn’t actually run a particle filter, but instead returns a function closure which does (see the previous post for a discussion of lexical scope and function closures in R). There are obviously several different steps within the particle filter, and several of these are amenable to parallelisation. However, for complex models, forward simulation from the model will be the rate-limiting step, where the vast majority of CPU cycles will be spent. Line 9 in the above code is where forward simulation takes place, and in particular, the key function call is the apply call:

apply(xmat, 1, stepFun, t0 = times[i], deltat = deltas[i], ...)

This call applies the forward simulation algorithm stepFun to each row of the matrix xmat independently. Since there are no dependencies between the function calls, this is in principle very straightforward to parallelise on multicore hardware.

Multicore support in R

I’m writing this post on a laptop with an Intel i7 quad core chip, running the 64 bit version of Ubuntu 11.10. R has support for multicore processing on this platform – it is just a simple matter of installing the relevant packages. However, things are changing rapidly regarding multicore support in R right now, so YMMV. Ubuntu 11.10 has R 2.13 by default, but the multicore support is slightly different in the recently released R 2.14. I’m still using R 2.13. I may update this post (or comment) when I move to R 2.14. The main difference is that the package multicore has been replaced by the package parallel. There are a few other minor changes, but it should be easy to adapt what is presented here to 2.14.

There is a new O’Reilly book called Parallel R. I’ve got a copy of it. It does cover the new parallel package in R 2.14, as well as other parallel R topics, but the book is a bit light weight, to say the least, and I reviewed it on this blog. Please read my review for further details before you buy it.

If you haven’t used multicore in R previously, then

install.packages(c("multicore","doMC"))

should get you started (again, I’m assuming that your R version is strictly < 2.14). You can test it has worked with:

library(multicore)
multicore:::detectCores()

When I do this, I get the answer 8 (I have 4 cores, each of which is hyper-threaded). To begin with, I want to tell R to use just 4 process threads, and I can do this with

library(doMC)
registerDoMC(4)

Replacing the second line with registerDoMC() will set things up to use all detected cores (in my case, 8). There are a couple of different strategies we could use to parallelise this. One strategy for parallelising the apply call discussed above is to be to replace it with a foreach / %dopar% loop. This is best illustrated by example. Start with line 9 from the function pfMLLik:

xmat = t(apply(xmat, 1, stepFun, t0 = times[i], deltat = deltas[i], ...))

We can produce a parallelised version by replacing this line with the following block of code:

res=foreach(j=1:dim(xmat)[1]) %dopar% {
  stepFun(xmat[j,], t0 = times[i], deltat = deltas[i], ...)
}
xmat=t(sapply(res,cbind))

Each iteration of the foreach loop is executed independently (possibly using multiple cores), and the result of each iteration is returned as a list, and captured in res. This list of return vectors is then coerced back into a matrix with the final line.

In fact, we can improve on this by using the .combine argument to foreach, which describes how to combine the results from each iteration. Here we can just use rbind to combine the results into a matrix, using:

xmat=foreach(j=1:dim(xmat)[1], .combine="rbind") %dopar% {
  stepFun(xmat[j,], t0 = times[i], deltat = deltas[i], ...)
}

This code is much neater, and in principle ought to be a bit faster, though I haven’t noticed much difference in practice.

In fact, it is not necessary to use the foreach construct at all. The multicore package provides the mclapply function, which is a multicore version of lapply. To use mclapply (or, indeed, lapply) here, we first need to split our matrix into a list of rows, which we can do using the split command. So in fact, our apply call can be replaced with the single line:

xmat=t(sapply(mclapply(split(xmat,row(xmat)), stepFun, t0=times[i], deltat=deltas[i], ...),cbind))

This is actually a much cleaner solution than the method using foreach, but it does require grokking a bit more R. Note that mclapply uses a different method to specify the number of threads to use than foreach/doMC. Here you can either use the named argument to mclapply, mc.cores, or use options(), eg. options(cores=4).

As well as being much cleaner, I find that the mclapply approach is much faster than the foreach/dopar approach for this problem. I’m guessing that this is because foreach doesn’t pre-schedule tasks by default, whereas mclapply does, but I haven’t had a chance to dig into this in detail yet.

A parallelised particle filter

We can now splice the parallelised forward simulation step (using mclapply) back into our particle filter function to get:

require(multicore)
pfMLLik <- function (n, simx0, t0, stepFun, dataLik, data) 
{
    times = c(t0, as.numeric(rownames(data)))
    deltas = diff(times)
    return(function(...) {
        xmat = simx0(n, t0, ...)
        ll = 0
        for (i in 1:length(deltas)) {
        xmat=t(sapply(mclapply(split(xmat,row(xmat)), stepFun, t0=times[i], deltat=deltas[i], ...),cbind))
            w = apply(xmat, 1, dataLik, t = times[i + 1], y = data[i,], log = FALSE, ...)
            if (max(w) < 1e-20) {
                warning("Particle filter bombed")
                return(-1e+99)
            }
            ll = ll + log(mean(w))
            rows = sample(1:n, n, replace = TRUE, prob = w)
            xmat = xmat[rows, ]
        }
        ll
    })
}

This can be used in place of the version supplied with the smfsb package for slow simulation algorithms running on modern multicore machines.

There is an issue regarding Monte Carlo simulations such as this and the multicore package (whether you use mclapply or foreach/dopar) in that it adopts a “different seeds” approach to parallel random number generation, rather than a true parallel random number generator. This probably isn’t worth worrying too much about now, since it is fixed in the new parallel package in R 2.14, but is something to be aware of. I discuss parallel random number generation issues in Wilkinson (2005).

Granularity

The above code is now a parallel particle filter, and can now be used in place of the serial version that is part of the smfsb package. However, if you try it out on a simple example, you will most likely be disappointed. In particular, if you use it for the pMCMC example discussed in the previous post, you will see that the parallel version of the example actually runs much slower than the serial version (at least, it does for me). However, that is because the forward simulator stepFun, used in that example, was actually a very fast simulation algorithm, stepLVc, written in C. In this case, the overhead of setting up and closing down the threads, and distributing the tasks, and collating the results from the worker threads back in the master thread, etc., outweighs the advantage of running the individual tasks in parallel. This is why parallel programming is difficult. What is needed here is for the individual tasks to be sufficiently computationally intensive that the overheads associated with parallelisation are easily outweighed by the ability to run the tasks in parallel. In the context of particle filtering, this is particularly problematic, as if the forward simulator is very slow, running a reasonable particle filter is going to be very, very slow, and then you probably don’t want to be working in R anyway… Parallelising a particle filter written in C using MPI is much more likely to be successful, as it offers much more fine grained control of exactly how the tasks and processes are managed, but at the cost of increased development time. In a previous post I gave an introduction to parallel Monte Carlo with C and MPI, and I’ve written more extensively about parallel MCMC in Wilkinson (2005). It also looks as though the new parallel package in R 2.14 offers more control of parallelisation, so that also might help. However, if you are using a particle filter as part of a pMCMC algorithm, there is another strategy you can use at a higher level of granularity which might be useful even within R in some situations.

Multiple particle filters and pMCMC

Let’s look again at the main loop of the pMCMC algorithm discussed in the previous post:

for (i in 1:iters) {
	message(paste(i,""),appendLF=FALSE)
	for (j in 1:thin) {
		thprop=th*exp(rnorm(p,0,tune))
		llprop=mLLik(thprop)
		if (log(runif(1)) < llprop - ll) {
			th=thprop
			ll=llprop
			}
		}
	thmat[i,]=th
	}

It is clear that the main computational bottleneck of this code is the call to mLLik on line 5, as this is the call which runs the particle filter. The purpose of making the call is to obtain an unbiased estimate of marginal likelihood. However, there are plenty of other ways that we can obtain such estimates than by running a single particle filter. In particular, we could run multiple particle filters and average the results. So, let’s look at how to do this in the multicore setting. Let’s start by thinking about running 4 particle filters. We could just replace the line

llprop=mLLik(thprop)

with the code

llprop=0.25*foreach(i=1:4, .combine="+") %dopar% {
    mLLik(thprop)
    }

Now, there are at least 2 issues with this. The first is that we are now just running 4 particle filters rather than 1, and so even with perfect parallelisation, it will run no quicker than the code we started with. However, the idea is that by running 4 particle filters we ought to be able to get away with each particle filter using fewer particles, though it isn’t trivial to figure out exactly how many. For example, averaging the results from 4 particle filters, each of which uses 25 particles is not as good as running a single particle filter with 100 particles. In practice, some trial and error is likely to be required. The second problem is that we have computed the mean of the log of the likelihoods, and not the likelihoods themselves. This will almost certainly work fine in practice, as the resulting estimate will in most cases be very close to unbiased, but it will not be exactly unbiased, as so will not lead to an “exact” approximate algorithm. In principle, this can be fixed by instead using

res=foreach(i=1:4) %dopar% {
    mLLik(thprop)
    }
llprop=log(mean(sapply(res,exp)))

but in practice this is likely to be subject to numerical underflow problems, as it involves manipulating raw likelihood values, which is generally a bad idea. It is possible to compute the log of the mean of the likelihoods in a more numerically stable way, but that is left as an exercise for the reader, as this post is way too long already… However, one additional tip worth mentioning is that the foreach package includes a convenience function called times for situations like the above, where the argument is not varying over calls. So the above code can be replaced with

res=times(4) %dopar% mLLik(thprop)
llprop=log(mean(sapply(res,exp)))

which is a bit cleaner and more readable.

Using this approach to parallelisation, there is now a much better chance of getting some speedup on multicore architectures, as the granularity of the tasks being parallelised is now much larger. Consider the example from the previous post, where at each iteration we ran a particle filter with 100 particles. If we now re-run that example, but instead use 4 particle filters each using 25 particles, we do get a slight speedup. However, on my laptop, the speedup is only around a factor of 1.6 using 4 cores, and as already discussed, 4 filters each with 25 particles isn’t actually quite as good as a single filter with 100 particles anyway. So, the benefits are rather modest here, but will be much better with less trivial examples (slower simulators). For completeness, a complete runnable demo script is included after the references. Also, it is probably worth emphasising that if your pMCMC algorithm has a short burn-in period, you may well get much better overall speed-ups by just running parallel MCMC chains. Depressing, perhaps, but true.

References

McCallum, E., Weston, S. (2011) Parallel R, O’Reilly.

Wilkinson, D. J. (2005) Parallel Bayesian Computation, Chapter 16 in E. J. Kontoghiorghes (ed.) Handbook of Parallel Computing and Statistics, Marcel Dekker/CRC Press, 481-512.

Wilkinson, D. J. (2011) Stochastic Modelling for Systems Biology, second edition, Boca Raton, Florida: Chapman & Hall/CRC Press.

Demo script

require(smfsb)
data(LVdata)

require(multicore)
require(doMC)
registerDoMC(4)

# set up data likelihood
noiseSD=10
dataLik <- function(x,t,y,log=TRUE,...)
{
	ll=sum(dnorm(y,x,noiseSD,log=TRUE))
	if (log)
		return(ll)
	else
		return(exp(ll))
}
# now define a sampler for the prior on the initial state
simx0 <- function(N,t0,...)
{
	mat=cbind(rpois(N,50),rpois(N,100))
	colnames(mat)=c("x1","x2")
	mat
}
# convert the time series to a timed data matrix
LVdata=as.timedData(LVnoise10)
# create marginal log-likelihood functions, based on a particle filter

# use 25 particles instead of 100
mLLik=pfMLLik(25,simx0,0,stepLVc,dataLik,LVdata)

iters=1000
tune=0.01
thin=10
th=c(th1 = 1, th2 = 0.005, th3 = 0.6)
p=length(th)
ll=-1e99
thmat=matrix(0,nrow=iters,ncol=p)
colnames(thmat)=names(th)
# Main pMCMC loop
for (i in 1:iters) {
	message(paste(i,""),appendLF=FALSE)
	for (j in 1:thin) {
		thprop=th*exp(rnorm(p,0,tune))
		res=times(4) %dopar% mLLik(thprop)
		llprop=log(mean(sapply(res,exp)))
		if (log(runif(1)) < llprop - ll) {
			th=thprop
			ll=llprop
			}
		}
	thmat[i,]=th
	}
message("Done!")
# Compute and plot some basic summaries
mcmcSummary(thmat)

Particle filtering and pMCMC using R

In the previous post I gave a quick introduction to the CRAN R package smfsb, and how it can be used for simulation of Markov processes determined by stochastic kinetic networks. In this post I’ll show how to use data and particle MCMC techniques in order to carry out Bayesian inference for the parameters of partially observed Markov processes.

The simulation model and the data

For this post we will assume that the smfsb package is installed and loaded (see previous post for details). The package includes the function stepLVc which simulates from the Markov transition kernel of a Lotka-Volterra process by calling out to some native C code for speed. So, for example,

stepLVc(c(x1=50,x2=100),0,1)

will simulate the state of the process at time 1 given an initial condition of 50 prey and 100 predators at time 0, using the default rate parameters of the function, th = c(1, 0.005, 0.6). The package also includes some data simulated from this model using these parameters, with and without added noise. The datasets can be loaded with

data(LVdata)

For simplicity, we will just make use of the dataset LVnoise10 in this post. This dataset is a multivariate time series consisting of 16 equally spaced observations on both prey and predators subject to Gaussian measurement error with a standard deviation of 10. We can plot the data with

plot(LVnoise10,plot.type="single",col=c(2,4))

giving:

The Bayesian inference problem is to see how much we are able to learn about the parameters which generated the data using only the data and our knowledge of the structure of the problem. There are many approaches one can take to this problem, but most are computationally intensive, due to the analytical intractability of the transition kernel of the LV process. Here we will follow Wilkinson (2011) and take a particle MCMC (pMCMC) approach, and specifically, use a pseudo-marginal “exact approximate” MCMC algorithm based on the particle marginal Metropolis-Hastings (PMMH) algorithm. I have discussed the pseudo-marginal approach, using particle filters for marginal likelihood estimation, and the PMMH algorithm in previous posts, so if you have been following my posts for a while, this should all make perfect sense…

Particle filter

One of the key ingredients required to implement the pseudo-marginal MCMC scheme is a (bootstrap) particle filter which generates an unbiased estimate of the marginal likelihood of the data given the parameters (integrated over the unobserved state trajectory). The algorithm was discussed in this post, and R code to implement this is included in the smfsb R package as pfMLLik. For reasons of numerical stability, the function computes and returns the log of the marginal likelihood, but it is important to understand that it is the actually likelihood estimate that is unbiased for the true likelihood, and not the corresponding statement for the logs. The actual code of the function is relatively short, and for completeness is given below:

pfMLLik <- function (n, simx0, t0, stepFun, dataLik, data) 
{
    times = c(t0, as.numeric(rownames(data)))
    deltas = diff(times)
    return(function(...) {
        xmat = simx0(n, t0, ...)
        ll = 0
        for (i in 1:length(deltas)) {
            xmat = t(apply(xmat, 1, stepFun, t0 = times[i], deltat = deltas[i], ...))
            w = apply(xmat, 1, dataLik, t = times[i + 1], y = data[i,], log = FALSE, ...)
            if (max(w) < 1e-20) {
                warning("Particle filter bombed")
                return(-1e+99)
            }
            ll = ll + log(mean(w))
            rows = sample(1:n, n, replace = TRUE, prob = w)
            xmat = xmat[rows, ]
        }
        ll
    })
}

We need to set up the prior and the data likelihood correctly before we can use this function, but first note that the function does not actually run a particle filter at all, but instead stores everything it needs to know to run the particle filter in the local environment, and then returns a function closure for evaluating the marginal likelihood at a given set of parameters. The resulting function (closure) can then be used to run a particle filter for a given set of parameters, simply by passing the required parameters into the function. This functional programming style is consistent with that used throughout the smfsb R package, and leads to quite simple, modular code. To use pfMLLik, we first need to define a function which evaluates the log-likelihood of an observation conditional on the true state, and another which samples from the prior distribution of the initial state of the system. Here, we can do that as follows.

# set up data likelihood
noiseSD=10
dataLik <- function(x,t,y,log=TRUE,...)
{
	ll=sum(dnorm(y,x,noiseSD,log=TRUE))
	if (log)
		return(ll)
	else
		return(exp(ll))
}
# now define a sampler for the prior on the initial state
simx0 <- function(N,t0,...)
{
	mat=cbind(rpois(N,50),rpois(N,100))
	colnames(mat)=c("x1","x2")
	mat
}
# convert the time series to a timed data matrix
LVdata=as.timedData(LVnoise10)
# create marginal log-likelihood functions, based on a particle filter
mLLik=pfMLLik(100,simx0,0,stepLVc,dataLik,LVdata)

Now the function (closure) mLLik will, for a given parameter vector, run a particle filter (using 100 particles) and return the log of the particle filter’s unbiased estimate of the marginal likelihood of the data. It is then very easy to use this function to create a simple PMMH algorithm for parameter inference.

PMMH algorithm

Below is an algorithm based on flat priors and a simple Metropolis-Hastings update for the parameters using the function closure mLLik, defined above.

iters=1000
tune=0.01
thin=10
th=c(th1 = 1, th2 = 0.005, th3 = 0.6)
p=length(th)
ll=-1e99
thmat=matrix(0,nrow=iters,ncol=p)
colnames(thmat)=names(th)
# Main pMCMC loop
for (i in 1:iters) {
	message(paste(i,""),appendLF=FALSE)
	for (j in 1:thin) {
		thprop=th*exp(rnorm(p,0,tune))
		llprop=mLLik(thprop)
		if (log(runif(1)) < llprop - ll) {
			th=thprop
			ll=llprop
			}
		}
	thmat[i,]=th
	}
message("Done!")
# Compute and plot some basic summaries
mcmcSummary(thmat)

This will take a little while to run, but in the end should give a plot something like the following (click for full size):

So, although we should really run the chain for a bit longer, we see that we can learn a great deal about the parameters of the process from very little data. For completeness, a full runnable demo script is included below the references. Of course there are many obvious extensions of this basic problem, such as partial observation (eg. only observing the prey) and unknown measurement error. These are discussed in Wilkinson (2011), and code for these cases is included within the demo(PMCMC), which should be inspected for further details.

Discussion

At this point it is probably worth emphasising that there are other “likelihood free” approaches which can be taken to parameter inference for partially observed Markov process (POMP) models. Many of these are implemented in the pomp R package, also available from CRAN, by King et al (2008). The pomp package is well documented, and has a couple of good tutorial vignettes which should be sufficient to get people started. The API of the package is rather cumbersome, but the algorithms appear to be quite robust. Approximate Bayesian computation (ABC) approaches are also quite natural for POMP models (see, for example, Toni et al (2009)). This is because “exact” likelihood free procedures break down in the case of low/no measurement error or high-dimensional observations. There are some R packages for ABC, but I am not sufficiently familiar with them to be able to give recommendations.

If one is able to move away from the “likelihood free” paradigm, it is possible to develop “exact” pMCMC algorithms which do not break down in challenging observation scenarios. The problem here is the intractability of the Markov transition kernel. In the case of nonlinear Markov jump processes, finding very generic solutions seems quite difficult, but for diffusion (approximation) processes based on stochastic differential equations, it seems to be possible to develop irreducible pMCMC algorithms which have very broad applicability – see Golightly and Wilkinson (2011) for further details of how such techniques can be used in the context of stochastic kinetic models similar to those considered in this post.

References

Golightly, A., Wilkinson, D. J. (2011) Bayesian parameter inference for stochastic biochemical network models using particle MCMC, Interface Focus, 1(6):807-820.

King, A.A., Ionides, E.L., & Breto, C.M. (2008) pomp: Statistical inference for partially observed Markov processes, CRAN.

Toni, T., Welch, D., Strelkowa, N., Ipsen, A. & Stumpf, M. (2009) Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems, J. R. Soc. Interface 6(31): 187-202.

Wilkinson, D. J. (2011) Stochastic Modelling for Systems Biology, second edition, Boca Raton, Florida: Chapman & Hall/CRC Press.

Demo script

require(smfsb)
data(LVdata)

# set up data likelihood
noiseSD=10
dataLik <- function(x,t,y,log=TRUE,...)
{
	ll=sum(dnorm(y,x,noiseSD,log=TRUE))
	if (log)
		return(ll)
	else
		return(exp(ll))
}
# now define a sampler for the prior on the initial state
simx0 <- function(N,t0,...)
{
	mat=cbind(rpois(N,50),rpois(N,100))
	colnames(mat)=c("x1","x2")
	mat
}
# convert the time series to a timed data matrix
LVdata=as.timedData(LVnoise10)
# create marginal log-likelihood functions, based on a particle filter
mLLik=pfMLLik(100,simx0,0,stepLVc,dataLik,LVdata)

iters=1000
tune=0.01
thin=10
th=c(th1 = 1, th2 = 0.005, th3 = 0.6)
p=length(th)
ll=-1e99
thmat=matrix(0,nrow=iters,ncol=p)
colnames(thmat)=names(th)
# Main pMCMC loop
for (i in 1:iters) {
	message(paste(i,""),appendLF=FALSE)
	for (j in 1:thin) {
		thprop=th*exp(rnorm(p,0,tune))
		llprop=mLLik(thprop)
		if (log(runif(1)) < llprop - ll) {
			th=thprop
			ll=llprop
			}
		}
	thmat[i,]=th
	}
message("Done!")
# Compute and plot some basic summaries
mcmcSummary(thmat)

The particle marginal Metropolis-Hastings (PMMH) particle MCMC algorithm

Introduction

In the previous post I explained how one can use an unbiased estimate of marginal likelihood derived from a particle filter within a Metropolis-Hastings MCMC algorithm in order to construct an exact pseudo-marginal MCMC scheme for the posterior distribution of the model parameters given some time course data. This idea is closely related to that of the particle marginal Metropolis-Hastings (PMMH) algorithm of Andreiu et al (2010), but not really exactly the same. This is because for a Bayesian model with parameters , latent variables and data , of the form

the pseudo-marginal algorithm which exploits the fact that the particle filter’s estimate of likelihood is unbiased is an MCMC algorithm which directly targets the marginal posterior distribution . On the other hand, the PMMH algorithm is an MCMC algorithm which targets the full joint posterior distribution . Now, the PMMH scheme does reduce to the pseudo-marginal scheme if samples of are not generated and stored in the state of the Markov chain, and it certainly is the case that the pseudo-marginal algorithm gives some insight into why the PMMH algorithm works. However, the PMMH algorithm is much more powerful, as it solves the “smoothing” and parameter estimation problem simultaneously and exactly, including the “initial value” problem (computing the posterior distribution of the initial state, ). Below I will describe the algorithm and explain why it works, but first it is necessary to understand the relationship between marginal, joint and “likelihood-free” MCMC updating schemes for such latent variable models.

MCMC for latent variable models

Marginal approach

If we want to target directly, we can use a Metropolis-Hastings scheme with a fairly arbitrary proposal distribution for exploring , where a new is proposed from and accepted with probability , where

As previously discussed, the problem with this scheme is that the marginal likelihood required in the acceptance ratio is often difficult to compute.

Likelihood-free MCMC

A simple “likelihood-free” scheme targets the full joint posterior distribution . It works by exploiting the fact that we can often simulate from the model for the latent variables even when we can’t evaluate it, or marginalise out of the problem. Here the Metropolis-Hastings proposal is constructed in two stages. First, a proposed new is sampled from and then a corresponding is simulated from the model . The pair is then jointly accepted with ratio

The proposal mechanism ensures that the proposed is consistent with the proposed , and so the procedure can work provided that the dimension of the data is low. However, in order to work well more generally, we would want the proposed latent variables to be consistent with the data as well as the model parameters.

Ideal joint update

Motivated by the likelihood-free scheme, we would really like to target the joint posterior by first proposing from and then a corresponding from the conditional distribution . The pair is then jointly accepted with ratio

Notice how the acceptance ratio simplifies, using the basic marginal likelihood identity (BMI) of Chib (1995), and drops out of the ratio completely in order to give exactly the ratio used for the marginal updating scheme. Thus, the “ideal” joint updating scheme reduces to the marginal updating scheme if is not sampled and stored as a component of the Markov chain.

Understanding the relationship between these schemes is useful for understanding the PMMH algorithm. Indeed, we will see that the “ideal” joint updating scheme (and the marginal scheme) corresponds to PMMH using infinitely many particles in the particle filter, and that the likelihood-free scheme corresponds to PMMH using exactly one particle in the particle filter. For an intermediate number of particles, the PMMH scheme is a compromise between the “ideal” scheme and the “blind” likelihood-free scheme, but is always likelihood-free (when used with a bootstrap particle filter) and always has an acceptance ratio leaving the exact posterior invariant.

The PMMH algorithm

The algorithm

The PMMH algorithm is an MCMC algorithm for state space models jointly updating and , as the algorithms above. First, a proposed new is generated from a proposal , and then a corresponding is generated by running a bootstrap particle filter (as described in the previous post, and below) using the proposed new model parameters, , and selecting a single trajectory by sampling once from the final set of particles using the final set of weights. This proposed pair is accepted using the Metropolis-Hastings ratio

where is the particle filter’s (unbiased) estimate of marginal likelihood, described in the previous post, and below. Note that this approach tends to the perfect joint/marginal updating scheme as the number of particles used in the filter tends to infinity. Note also that for a single particle, the particle filter just blindly forward simulates from and that the filter’s estimate of marginal likelihood is just the observed data likelihood leading precisely to the simple likelihood-free scheme. To understand for an arbitrary finite number of particles, , one needs to think carefully about the structure of the particle filter.

Why it works

To understand why PMMH works, it is necessary to think about the joint distribution of all random variables used in the bootstrap particle filter. To this end, it is helpful to re-visit the particle filter, thinking carefully about the resampling and propagation steps.

First introduce notation for the “particle cloud”: , , . Initialise the particle filter with , where and (note that is undefined). Now suppose at time we have a sample from : . First resample by sampling , . Here we use for the discrete distribution on with probability mass function . Next sample . Set and . Finally, propagate to the next step… We define the filter’s estimate of likelihood as and . See Doucet et al (2001) for further theoretical background on particle filters and SMC more generally.

Describing the filter carefully as above allows us to write down the joint density of all random variables in the filter as

For PMMH we also sample a final index from giving the joint density

We write the final selected trajectory as

where , and . If we now think about the structure of the PMMH algorithm, our proposal on the space of all random variables in the problem is in fact

and by considering the proposal and the acceptance ratio, it is clear that detailed balance for the chain is satisfied by the target with density proportional to

We want to show that this target marginalises down to the correct posterior when we consider just the parameters and the selected trajectory. But if we consider the terms in the joint distribution of the proposal corresponding to the trajectory selected by , this is given by

which, by expanding the in terms of the unnormalised weights, simplifies to

It is worth dwelling on this result, as this is the key insight required to understand why the PMMH algorithm works. The whole point is that the terms in the joint density of the proposal corresponding to the selected trajectory exactly represent the required joint distribution modulo a couple of normalising constants, one of which is the particle filter’s estimate of marginal likelihood. Thus, by including in the acceptance ratio, we knock out the normalising constant, allowing all of the other terms in the proposal to be marginalised away. In other words, the target of the chain can be written as proportional to

The other terms are all probabilities of random variables which do not occur elsewhere in the target, and hence can all be marginalised away to leave the correct posterior

Thus the PMMH algorithm targets the correct posterior for any number of particles, . Also note the implied uniform distribution on the selected indices in the target.

I will give some code examples in a future post.

MCMC, Monte Carlo likelihood estimation, and the bootstrap particle filter

The pseudo-marginal approach to MCMC for Bayesian inference

In a previous post I described a generalisation of the Metropolis Hastings MCMC algorithm which uses unbiased Monte Carlo estimates of likelihood in the acceptance ratio, but is nevertheless exact, when considered as a pseudo-marginal approach to “exact approximate” MCMC. To be useful in the context of Bayesian inference, we need to be able to compute unbiased estimates of the (marginal) likelihood of the data given some proposed model parameters with any “latent variables” integrated out.

To be more precise, consider a model for data with parameters of the form together with a prior on , , giving a joint model

Suppose now that interest is in the posterior distribution

We can construct a fairly generic (marginal) MCMC scheme for this posterior by first proposing from some fairly arbitrary proposal distribution and then accepting the value with probability where

This method is great provided that the (marginal) likelihood of the data is available to us analytically, but in many (most) interesting models it is not. However, in the previous post I explained why substituting in a Monte Carlo estimate will still lead to the exact posterior if the estimate is unbiased in the sense that . Consequently, sources of (cheap) unbiased Monte Carlo estimates of (marginal) likelihood are of potential interest in the development of exact MCMC algorithms.

Latent variables and marginalisation

Often the reason that we cannot evaluate is that there are latent variables in the problem, and the model for the data is conditional on those latent variables. Explicitly, if we denote the latent variables by , then the joint distribution for the model takes the form

Now since

there is a simple and obvious Monte Carlo strategy for estimating provided that we can evaluate and simulate realisations from . That is, simulate values from for some suitably large , and then put

It is clear by the law of large numbers that this estimate will converge to as . That is, is a consistent estimate of . However, a moment’s thought reveals that this estimate is not only consistent, but also unbiased, since each term in the sum has expectation . This simple Monte Carlo estimate of likelihood can therefore be substituted into a Metropolis-Hastings acceptance ratio without affecting the (marginal) target distribution of the Markov chain. Note that this estimate of marginal likelihood is sometimes referred to as the Rao-Blackwellised estimate, due to its connection with the Rao-Blackwell theorem.

Importance sampling

Suppose now that we cannot sample values directly from , but can sample instead from a distribution having the same support as . We can then instead produce an importance sampling estimate for by noting that

Consequently, samples from can be used to construct the estimate

which again is clearly both consistent and unbiased. This estimate is often written

where . The weights, , are known as importance weights.

Importance resampling

An idea closely related to that of importance sampling is that of importance resampling where importance weights are used to resample a sample in order to equalise the weights, often prior to a further round of weighting and resampling. The basic idea is to generate an approximate sample from a target density using values sampled from an auxiliary distribution , where we now supress any dependence of the distributions on model parameters, .

First generate a sample from and compute weights . Then compute normalised weights . Generate a new sample of size by sampling times with replacement from the original sample with the probability of choosing each value determined by its normalised weight.

As an example, consider using a sample from the Cauchy distribution as an auxiliary distribution for approximately sampling standard normal random quantities. We can do this using a few lines of R as follows.

n=1000
xa=rcauchy(n)
w=dnorm(xa)/dcauchy(xa)
x=sample(xa,n,prob=w,replace=TRUE)
hist(x,30)
mean(w)

Note that we don’t actually need to compute the normalised weights, as the sample function will do this for us. Note also that the average weight will be close to one. It should be clear that the expected value of the weights will be exactly 1 when both the target and auxiliary densities are correctly normalised. Also note that the procedure can be used when one or both of the densities are not correctly normalised, since the weights will be normalised prior to sampling anyway. Note that in this case the expected weight will be the (ratio of) normalising constant(s), and so looking at the average weight will give an estimate of the normalising constant.

Note that the importance sampling procedure is approximate. Unlike a technique such as rejection sampling, which leads to samples having exactly the correct distribution, this is not the case here. Indeed, it is clear that in the case, the final sample will be exactly drawn from the auxiliary and not the target. The procedure is asymptotic, in that it improves as the sample size increases, tending to the exact target as .

We can understand why importance resampling works by first considering the univariate case, using correctly normalised densities. Consider a very large number of particles, . The proportion of the auxiliary samples falling in a small interval will be , corresponding to roughly particles. The weight for each of those particles will be , and since the expected weight of a random particle is 1, the sum of all weights will be (roughly) , leading to normalised weights for the particles near of . The combined weight of all particles in is therefore . Clearly then, when we resample times we expect to select roughly particles from this interval. This corresponds to a proportion , corresponding to a density of in the final sample.

Obviously the above argument is very informal, but can be tightened up into a reasonably rigorous proof for the 1d case without too much effort, and the multivariate extension is also reasonably clear.

The bootstrap particle filter

The bootstrap particle filter is an iterative method for carrying out Bayesian inference for dynamic state space (partially observed Markov process) models, sometimes also known as hidden Markov models (HMMs). Here, an unobserved Markov process, governed by a transition kernel is partially observed via some measurement model leading to data . The idea is to make inference for the hidden states given the data . The method is a very simple application of the importance resampling technique. At each time, , we assume that we have a (approximate) sample from and use importance resampling to generate an approximate sample from .

More precisely, the procedure is initialised with a sample from with uniform normalised weights . Then suppose that we have a weighted sample from . First generate an equally weighted sample by resampling with replacement times to obtain (giving an approximate random sample from ). Note that each sample is independently drawn from . Next propagate each particle forward according to the Markov process model by sampling (giving an approximate random sample from ). Then for each of the new particles, compute a weight , and then a normalised weight .

It is clear from our understanding of importance resampling that these weights are appropriate for representing a sample from , and so the particles and weights can be propagated forward to the next time point. It is also clear that the average weight at each time gives an estimate of the marginal likelihood of the current data point given the data so far. So we define

and

Again, from our understanding of importance resampling, it should be reasonably clear that is a consistent estimator of . It is much less clear, but nevertheless true that this estimator is also unbiased. The standard reference for this fact is Del Moral (2004), but this is a rather technical monograph. A much more accessible proof (for a very general particle filter) is given in Pitt et al (2011).

It should therefore be clear that if one is interested in developing MCMC algorithms for state space models, one can use a pseudo-marginal MCMC scheme, substituting in from a bootstrap particle filter in place of . This turns out to be a simple special case of the particle marginal Metropolis-Hastings (PMMH) algorithm described in Andreiu et al (2010). However, the PMMH algorithm in fact has the full joint posterior as its target. I will explain the PMMH algorithm in a subsequent post.

The pseudo-marginal approach to “exact approximate” MCMC algorithms

Motivation and background

In this post I will try and explain an important idea behind some recent developments in MCMC theory. First, let me give some motivation. Suppose you are trying to implement a Metropolis-Hastings algorithm, as discussed in a previous post (required reading!), but a key likelihood term needed for the acceptance ratio is difficult to evaluate (most likely it is a marginal likelihood of some sort). If it is possible to obtain a Monte-Carlo estimate for that likelihood term (which it sometimes is, for example, using importance sampling), one could obviously just plug it in to the acceptance ratio and hope for the best. What is not at all obvious is that if your Monte-Carlo estimate satisfies some fairly weak property then the equilibrium distribution of the Markov chain will remain exactly as it would be if the exact likelihood had been available. Furthermore, it is exact even if the Monte-Carlo estimate is very noisy and imprecise (though the mixing of the chain will be poorer in this case). This is the “exact approximate” pseudo-marginal MCMC approach. To give credit where it is due, the idea was first introduced by Mark Beaumont in Beaumont (2003), where he was using an importance sampling based approximate likelihood in the context of a statistical genetics example. This was later picked up by Christophe Andrieu and Gareth Roberts, who studied the technical properties of the approach in Andrieu and Roberts (2009). The idea is turning out to be useful in several contexts, and in particular, underpins the interesting new Particle MCMC algorithms of Andrieu et al (2010), which I will discuss in a future post. I’ve heard Mark, Christophe, Gareth and probably others present this concept, but this post is most strongly inspired by a talk that Christophe gave at the IMS 2010 meeting in Gothenburg this summer.

The pseudo-marginal Metropolis-Hastings algorithm

Let’s focus on the simplest version of the problem, where we want to sample from a target using a proposal . As explained previously, the required Metropolis-Hastings acceptance ratio will take the form

Here we are assuming that is difficult to evaluate (usually because it is a marginalised version of some higher-dimensional distribution), but that a Monte-Carlo estimate of , which we shall denote , can be computed. We can obviously just substitute this estimate into the acceptance ratio to get

but it is not immediately clear that in many cases this will lead to the Markov chain having an equilibrium distribution that is exactly . It turns out that it is sufficient that the likelihood estimate, is non-negative, and unbiased, in the sense that , where the expectation is with respect to the Monte-Carlo error for a given fixed value of . In fact, as we shall see, this condition is actually a bit stronger than is really required.

Put , representing the noise in the Monte-Carlo estimate of , and suppose that (note that in an abuse of notation, the function is unrelated to ). The main condition we will assume is that , where is a constant independent of . In the case of , we have the (typical) special case of . For now, we will also assume that , but we will consider relaxing this constraint later.

The key to understanding the pseudo-marginal approach is to realise that at each iteration of the MCMC algorithm a new value of is being proposed in addition to a new value for . If we regard the proposal mechanism as a joint update of and , it is clear that the proposal generates from the density , and we can re-write our “approximate” acceptance ratio as

Inspection of this acceptance ratio reveals that the target of the chain must be (proportional to)

This is a joint density for , but the marginal for can be obtained by integrating over the range of with respect to . Using the fact that , this then clearly gives a density proportional to , and this is precisely the target that we require. Note that for this to work, we must keep the old value of from one iteration to the next – that is, we must keep and re-use our noisy value to include in the acceptance ratio for our next MCMC iteration – we should not compute a new Monte-Carlo estimate for the likelihood of the old state of the chain.

Examples

We will consider again the example from the previous post – simulation of a chain with a N(0,1) target using uniform innovations. Using R, the main MCMC loop takes the form

pmmcmc<-function(n=1000,alpha=0.5) 
{
        vec=vector("numeric", n)
        x=0
        oldlik=noisydnorm(x)
        vec[1]=x
        for (i in 2:n) {
                innov=runif(1,-alpha,alpha)
                can=x+innov
                lik=noisydnorm(can)
                aprob=lik/oldlik
                u=runif(1)
                if (u < aprob) { 
                        x=can
                        oldlik=lik
			}
                vec[i]=x
        }
        vec
}

Here we are assuming that we are unable to compute dnorm exactly, but instead only a Monte-Carlo estimate called noisydnorm. We can start with the following implementation

noisydnorm<-function(z)
{
	dnorm(z)*rexp(1,1)
}

Each time this function is called, it will return a non-negative random quantity whose expectation is dnorm(z). We can now run this code as follows.

plot.mcmc<-function(mcmc.out)
{
	op=par(mfrow=c(2,2))
	plot(ts(mcmc.out),col=2)
	hist(mcmc.out,30,col=3)
	qqnorm(mcmc.out,col=4)
	abline(0,1,col=2)
	acf(mcmc.out,col=2,lag.max=100)
	par(op)
}

metrop.out<-pmmcmc(10000,1)
plot.mcmc(metrop.out)

MCMC output and convergence diagnostics

So we see that we have exactly the right N(0,1) target, despite using the (very) noisy noisydnorm function in place of the dnorm function. This noisy likelihood function is clearly unbiased. However, as already discussed, a constant bias in the noise is also acceptable, as the following function shows.

noisydnorm<-function(z)
{
	dnorm(z)*rexp(1,2)
}

Re-running the code with this function also leads to the correct equilibrium distribution for the chain. However, it really does matter that the bias is independent of the state of the chain, as the following function shows.

noisydnorm<-function(z)
{
	dnorm(z)*rexp(1,0.1+10*z*z)
}

Running with this function leads to the wrong equilibrium. However, it is OK for the distribution of the noise to depend on the state, as long as its expectation does not. The following function illustrates this.

noisydnorm<-function(z)
{
	dnorm(z)*rgamma(1,0.1+10*z*z,0.1+10*z*z)
}

This works just fine. So far we have been assuming that our noisy likelihood estimates are non-negative. This is clearly desirable, as otherwise we could wind up with negative Metropolis-Hasting ratios. However, as long as we are careful about exactly what we mean, even this non-negativity condition may be relaxed. The following function illustrates the point.

noisydnorm<-function(z)
{
	dnorm(z)*rnorm(1,1)
}

Despite the fact that this function will often produce negative values, the equilibrium distribution of the chain still seems to be correct! An even more spectacular example follows.

noisydnorm<-function(z)
{
	dnorm(z)*rnorm(1,0)
}

Astonishingly, this one works too, despite having an expected value of zero! However, the following doesn’t work, even though it too has a constant expectation of zero.

noisydnorm<-function(z)
{
	dnorm(z)*rnorm(1,0,0.1+10*z*z)
}

I don’t have time to explain exactly what is going on here, so the precise details are left as an exercise for the reader. Suffice to say that the key requirement is that it is the distribution of W conditioned to be non-negative which must have the constant bias property – something clearly violated by the final noisydnorm example.

Before finishing this post, it is worth re-emphasising the issue of re-using old Monte-Carlo estimates. The following function will not work (exactly), though in the case of good Monte-Carlo estimates will often work tolerably well.

approxmcmc<-function(n=1000,alpha=0.5) 
{
        vec=vector("numeric", n)
        x=0
        vec[1]=x
        for (i in 2:n) {
                innov=runif(1,-alpha,alpha)
                can=x+innov
                lik=noisydnorm(can)
                oldlik=noisydnorm(x)
                aprob=lik/oldlik
                u=runif(1)
                if (u < aprob) { 
                        x=can
			}
                vec[i]=x
        }
        vec
}

In a subsequent post I will show how these ideas can be put into practice in the context of a Bayesian inference example.