Adaptive Strategy Selection within Differential Evolution on the BBOB-2010 Noiseless Benchmark

Álvaro Fialho (1) , Raymond Ros (2)

1 MSR - INRIA - Microsoft Research - Inria Joint Centre
2 TAO - Machine Learning and Optimisation

This document presents the adaptive strategy selection Fitness-Based Area-Under-Curve Bandit (F-AUC-Bandit) and its use in the context of a continuous optimization algorithm: Differential Evolution. Experimental results were obtained on a testbed of noiseless functions. They demonstrate the interest of using an adaptive strategy selection technique as opposed to the naïve approach which consists in the use of a unique strategy of the optimization algorithm chosen initially. Performance comparisons are made between F-AUC-Bandit and a uniform strategy selection approach and also other adaptive selection strategies. Finally a comparison to the state-of-the-art CMA-ES optimizer is made. Results of the optimization algorithm using F-AUC-Bandit are still not comparable to those of CMA-ES but demonstrate a big improvement on the use of the basic DE.

Domaines: Analyse numérique [cs.NA]; Apprentissage [cs.LG]

Álvaro Fialho, Raymond Ros. Adaptive Strategy Selection within Differential Evolution on the BBOB-2010 Noiseless Benchmark. [Research Report] RR-7259, 2010. ⟨inria-00476160v1⟩