All publications

IEO: Intelligent Evolutionary Optimisation for Hyperparameter Tuning

2020·arXiv
mutateevaluateselectbreedsurrogatepredicts fitnessskip predicted losersevaluationsbest scoreplain EA(a) learning inside the loop(b) faster convergence

Figure 1: Learning inside the loop. A surrogate model predicts which offspring merit real evaluation (a), shortcutting the evolutionary cycle without surrendering the final measurement (b).

Evolutionary hyperparameter search is robust but spendthrift: most evaluations are wasted on candidates a model could have predicted were poor. IEO puts a learned model inside the loop to steer it — predicting promising regions and skipping evaluations the surrogate is confident about.

On classification tasks this hybrid reached comparable configurations substantially faster than conventional evolutionary search — an early instance of the pattern that recurs through this research line: learning guides the search, measurement still decides.

Key results

  • Substantially faster hyperparameter search on classification tasks
  • Learned guidance reduces wasted evaluations
  • Search quality preserved under the accelerated loop