Skip to content

A parallel version of Loshchilov, Schoenauer and Sebag's (2012) adaptive coordinate descent algorithm for non-linear optimization

License

Notifications You must be signed in to change notification settings

tholden/ParallelFastNonLinearACD

Repository files navigation

ParallelFastNonLinearACD

A parallel version of Loshchilov, Schoenauer and Sebag's (2012) adaptive coordinate descent algorithm for non-linear optimization

Key reference

Loshchilov, I., Schoenauer, M. , Sebag, M. (2011). Adaptive Coordinate Descent.

in N. Krasnogor et al. (eds.), Genetic and Evolutionary Computation Conference (GECCO) 2012, Proceedings, ACM.

http://hal.inria.fr/docs/00/58/75/34/PDF/AdaptiveCoordinateDescent.pdf

Files

  • ACD0 is a slightly cleaned up version of the original code from here: https://uk.mathworks.com/matlabcentral/fileexchange/37057-fast-adaptive-coordinate-descent-for-non-linear-optimization
    • Extra features include support for linear constraints (with clipping to the bound) and slightly different input arguments (an initial point, and sigma, the initial search standard deviation(s)).
  • ACD is a (potentially) parallelized version of ACD0, with vectorization/parallelization controlled by the Order, NonProductSearchDimension and ProductSearchDimension inputs.
    • With NonProductSearchDimension * ProductSearchDimension = 1, the search is coordinate by coordinate.
      • At Order 1 this uses just 2 points per iteration, and will often be slower than serial code if the objective is parallel.
      • At Order 2 this uses 6 points per iteration, and may be faster than serial code when the serial objective function is slow to evaluate, and you have more than 6 cores.
      • At Order n this uses 2^(1+n)-2 points per iteration.
      • In all cases, the parallelization is providing improved univariate search in each iteration, reducing the required number of iterations.
    • With `NonProductSearchDimension * ProductSearchDimension = 2', the search is over pairs of coordinates, rather than coordinate by coordinate.
      • At Order 1, this uses 8 points per iteration if ProductSearchDimension = 2, or 6 points per iteration if NonProductSearchDimension = 2, so may be perfect for 8 core desktops in either case.
      • At Order n, this uses (2^(NonProductSearchDimension+n)-1)^ProductSearchDimension-1 points per iteration.
    • With NonProductSearchDimension * ProductSearchDimension = d, the search is over d-tuples of coordinates.
      • At Order n, this uses (2^(NonProductSearchDimension+n)-1)^ProductSearchDimension-1 points per iteration.

Function usage

[ xMean, BestFitness, Iterations, NEvaluations ] = ACD( FitnessFunction, xMean, sigma, LB, UB, A, b, MaxEvaluations, StopFitness, HowOftenUpdateRotation, Order, SearchDimension, Resume );

Inputs:

  • FitnessFunction: The objective function, a function handle. The objective must be vectorized, supporting a matrix of inputs (with one column per observation), and returning a vector of outputs. To assist with converting arbitrary functions to this form, three wrappers (SerialWrapper, ParForParallelWrapper, TimedParallelWrapper) are provided.
  • xMean: The initial point.
  • Sigma: The initial search radius. Either a scalar, or a vector of search radiuses by coordinate, with the same number of elements as xMean.
  • MinSigma: The minimum search radius. The search will stop when all coordinates of sigma are below this value. Either a scalar, or a vector of minimum search radiuses by coordinate, with the same number of elements as xMean.
  • LB: A lower bound on the search for xMean. Either empty, a scalar, or a vector of lower bounds by coordinate, with the same number of elements as xMean.
  • UB: A lower bound on the search for xMean. Either empty, a scalar, or a vector of upper bounds by coordinate, with the same number of elements as xMean.
  • A: The A matrix from the inequality A*x <= b. May be empty if b is also empty.
  • b: The b vector from the inequality A*x <= b. May be empty if b is also empty.
  • MaxEvaluations: The maximum number of total function evaluations. (Set to Inf if this is empty.)
  • StopFitness: The terminal fitness. (Set to -Inf if this is empty.)
  • HowOftenUpdateRotation: How often the rotation should be updated. On problems with slow objective functions, this should be equal to 1. Larger values may speed up computation if the objective function is very fast.
  • Order: Determines the number of points to use to search along each group of NonProductSearchDirection directions. A (small) positive integer.
  • NonProductSearchDimension: NonProductSearchDimension*ProductSearchDimension determines how many dimensions to search in simultaneously. A (small) positive integer.
  • ProductSearchDimension: NonProductSearchDimension*ProductSearchDimension determines how many dimensions to search in simultaneously. A (small) positive integer.
  • Resume: Whether to resume the past run. A logical.

Ouputs:

  • xMean: The optimal point.
  • BestFitness: The value of the objective at that point.
  • Iterations: The number of iterations performed.
  • NEvaluations: The number of function evaluations performed.

About

A parallel version of Loshchilov, Schoenauer and Sebag's (2012) adaptive coordinate descent algorithm for non-linear optimization

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published