Optimizing Algorithm Configuration by Improved Initial Heterogeneous Parameter Sampling

Irace [1] is a method for automatic configuration of stochastic optimization algorithms, also called hyper-parameter optimization. While the initial set of configurations tested by irace is obtained by uniform random sampling, other sampling strategies should also be possible. The goal of this project is to explore other initialization methods, including but not limited to:

  • Random sampling that takes the conditional structure of parameters into account
  • Latin Hyper-Cube Sampling
  • Exhaustive Parametrization
  • Low-discrepancy sequences: Halton, Hammersley set, Sobol, etc.

In some cases, a combination of methods may be necessary, in order to sample from numerical, categorical and ordinal parameters, where some parameters are conditional to others. Methods that are able to handle constraints on parameter values, either by resampling a posteriori or by adapting the sampling a priori, would also be of interest.

Literature / References:

  • [1] López-Ibáñez, M., Dubois-Lacoste, J., Pérez Cáceres, L., Stützle, T. and Birattari, M. (2016). The irace package: Iterated Racing for Automatic Algorithm Configuration. Operations Research Perspectives. [R-Package, pdf]