Visualization of the CMA-ES algorithm in a multi-modal function.
Derandomized Evolution Strategies and Local Learning
We develop Evolution Strategies (ES) for Optimization coupled with Local Learning models to tackle problems with expensive cost functions. ES are a class Bioinspired optimization Algorithms that mimic natural processes to find optimal designs. In the successful translation of the natural evolution process into effcient and robust computer algorithms, model building plays a central role. Local meta-models are used to replace costly evaluations of the objective function by cheap estimates. We investigate and enhance ESs and we apply them to challenging real world problems.
CMA-ES
The derandomized Evolution Strategy (ES) with Covariance Matrix Adaptation (CMA) adapts the complete covariance matrix of the normal mutation (search) distribution.
Sketch of the different steps characterizing the algorithm
Local meta-models
The effciency of EAs for expensive problems can be improved by incorporating local meta-models of the cost function. We enhance CMA-ES with a full quadratic local meta-model to improve the convergence speed of the CMA-ES.
Main features
- CMA-ES is as a robust local search strategy
- CMA-ES outperforms conventional optimization algorithms on problems that are discontinuous, non-differentiable, multi-modal and noisy
- CMA-ES efficiency is improved by the use of local meta-models
- It was successfully applied to a considerable number of real world problems
Applications
Given its robustness and efficiency, CMA-ES results particularly suitable to address parameter identification in real world problems. In fact CMA-ES has been successfully applied to applications ranging from virus and pedestrian traffic, to the study of anguilliform swimming.