Limited-Memory Matrix Adaptation for Large Scale Black-box Optimization
About
The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is a popular method to deal with nonconvex and/or stochastic optimization problems when the gradient information is not available. Being based on the CMA-ES, the recently proposed Matrix Adaptation Evolution Strategy (MA-ES) provides a rather surprising result that the covariance matrix and all associated operations (e.g., potentially unstable eigendecomposition) can be replaced in the CMA-ES by a updated transformation matrix without any loss of performance. In order to further simplify MA-ES and reduce its $\mathcal{O}\big(n^2\big)$ time and storage complexity to $\mathcal{O}\big(n\log(n)\big)$, we present the Limited-Memory Matrix Adaptation Evolution Strategy (LM-MA-ES) for efficient zeroth order large-scale optimization. The algorithm demonstrates state-of-the-art performance on a set of established large-scale benchmarks. We explore the algorithm on the problem of generating adversarial inputs for a (non-smooth) random forest classifier, demonstrating a surprising vulnerability of the classifier.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| High-dimensional optimization | MSLR | Convergence Value-8.6829 | 21 | |
| High-dimensional optimization | Lasso-Hard | Convergence Value5.5371 | 20 | |
| High-dimensional optimization | LIMO | Convergence Value-6.5043 | 20 | |
| Function Optimization | Rosenbrock D=1000 | Convergence Value4.48e+5 | 19 | |
| Function Optimization | Levy D=1000 | Convergence Value98.4986 | 19 | |
| Function Optimization | Sphere D=1000 | Final Value99.1157 | 19 | |
| Function Optimization | Dixon D=1000 | Convergence Value6.82e+5 | 19 | |
| Function Optimization | Griewank D=1000 | Convergence Value (Statistic)83.7417 | 19 | |
| Function Optimization | Michalewicz D=1000 | Convergence Value-7.3345 | 19 | |
| High-dimensional optimization | Rosenbrock D=10000 | Convergence Value5.72e+5 | 13 |