Covariance Matrix Adaptation Evolution Strategy

This library is a minimalistic Scala version of the CMA-ES algorithm as described in this paper. The implementation is not as full as Java version of the
algorithm however it provides certain benefits compared to the latter one:

• This implementation is more compact and straightforward.
• In Scala version objective function can be evaluated in parallel. This might help to reduce optimization time if it takes long time to compute your fitness function.
• The strategy implementation is stateless and idempotent.

Using CMA-ES

This project can be built with sbt 0.13. Add these lines to your SBT project definition:

resolvers += Resolver.bintrayRepo("volodymyr-orlov", "maven")

libraryDependencies ++= Seq(
  "com.sungevity" %% "cma-es-scala" % "1.0.2"
)

Here is a minimal working example.:

 
  import com.sungevity.cmaes._
  import breeze.linalg._

  val rosenbrock = ParallelObjectiveFunction ({ // a non-convex function we want to optimize
    case x: DenseVector[Double] => {
          (for (i <- 0 until x.length - 1) yield {
            100 * (x(i) * x(i) - x(i + 1)) * (x(i) * x(i) - x(i + 1)) +
              (x(i) - 1.0) * (x(i) - 1.0);
          }).sum
    }
  })
  
  val driver = CMAESDriver(rosenbrock) // instantiate a driver  
  
  val result = driver.optimize(5, // population size
    5, // search space size
    0.05, // initial value
    0.2, // initial standard deviation
    iterationsExceeded(6000) orElse lowVariance(1e-14) orElse minFitnessReached(1e-14) orElse proceed // stop condition
    )