JarCastingSchrodinger
Schrodinger is an (early-stage) project for probabilistic programming built on the Cats ecosystem. At its heart is the RandomT monad transformer which enables you to safely inject randomness into your program by describing it as a random variable while also interleaving with other effects, such as Cats Effect's IO monad. The goal is that this construct will facilitate the implementation of high-performance, concurrent Monte Carlo algorithms for simulation and inference.
Notably, via the use of "splittable" RNGs (versus a synchronized global or thread-locals), every random number generated can be deterministically traced back to the initial seed irrespective of execution order or thread assignment at runtime. This substantially improves the reproducibility of concurrent programs.
Usage
libraryDependencies += "com.armanbilge" %% "schrodinger" % "0.2.0"
// (optional) provides additional instances for cats-effect typeclasses
libraryDependencies += "com.armanbilge" %% "schrodinger-effect" % "0.2.0"
Datatypes
-
schrodinger.RandomT: a monad transformer for a random variable that describes a probabilistic program. It is a light wrapper for Cat'sStateTwith its own typeclass instances. Hints:- (Flat)Map the "primitive" random variables provided in
schrodinger.random.*to createRandomTinstances for your own datatypes. - Use typeclasses to defer binding the RNG state type
Sto a specific implementation until the (terminal) simulation step.
See this in action in the example.
- (Flat)Map the "primitive" random variables provided in
-
schrodinger.data.Weighted: represents a "weighted" sample from a probability distribution. It is similar to Cat'sWriterdatatype with the additional feature that it short-circuits computation (likeOption) if the weight goes to zero. This is useful for (nested) importance sampling where samples from an auxiliary distribution are reweighted to sample from a target distribution.
Example: Schrödinger's cat
In the example project we reimagine the famous thought experiment with the help of Cats Effect. The experiment is modeled with three concurrent components:
- an atom that decays after a random exponentially-distributed time and is sensed by the Geiger counter
- a relay that is triggered by the Geiger counter to release poison and kill the cat :(
- an observer who starts the experiment, waits 1 second, and then checks on the cat
Although the only randomness in the program is the atom's decay, we build the entire program as a RandomT[F[_], S, *] and defer simulating it until "right before the end of the world" at which point we provide the initial seed for the RNG.
sealed trait Cat
case object LiveCat extends Cat
case object DeadCat extends Cat
val decayRate = math.log(2)
def decayingAtom[F[_]: Async, S: SplitMonadCancel](
geigerCounter: CountDownLatch[RandomT[F, S, *]])(implicit E: ExponentialDouble[F, S]) =
for {
decayAfter <- Exponential[F, S](decayRate)
_ <- Async[RandomT[F, S, *]].sleep(decayAfter.seconds)
_ <- geigerCounter.release
} yield ()
def poisonRelay[F[_]: Monad](geigerCounter: CountDownLatch[F], cat: Ref[F, Cat]) =
for {
_ <- geigerCounter.await
_ <- cat.set(DeadCat)
} yield ()
def experiment[F[_]: Async, S: SplitMonadCancel](implicit E: ExponentialDouble[F, S]) =
for {
cat <- Ref.of[RandomT[F, S, *], Cat](LiveCat)
geigerCounter <- CountDownLatch[RandomT[F, S, *]](1)
// spawning fibers splits the RNG deterministically
_ <- poisonRelay[RandomT[F, S, *]](geigerCounter, cat).start
_ <- decayingAtom[F, S](geigerCounter).start
_ <- Async[RandomT[F, S, *]].sleep(1.second)
// 50% probability that we will observe a live cat
observation <- cat.get
} yield observation
val seed1 = SplitMix.initialState(0x2b992ddfa23249d6L, 0x4034650f1c98bd69L)
val seed2 = SplitMix.initialState(0x86d98163ff1fe751L, 0x8316a8fe31a2228eL)
override def run: IO[Unit] = for {
observation1 <- experiment[IO, SplitMix].simulate(seed1)
_ <- IO.println(s"Experiment 1: observing a $observation1")
observation2 <- experiment[IO, SplitMix].simulate(seed2)
_ <- IO.println(s"Experiment 2: observing a $observation2")
_ <- IO.println("No cats were harmed in the thinking of this experiment :)")
} yield ()
Run from the command line with sbt "example/runMain schrodinger.ThoughtExperiment". In spite of its concurrency, this program is virtually guaranteed to produce:
Experiment 1: observing a LiveCat
Experiment 2: observing a DeadCat
No cats were harmed in the thinking of this experiment :)