Natural evolution strategies converge on sphere functions

Authors:
Tom Schaul
Affiliations:
New York University, New York, NY, USA
Venue:
Proceedings of the 14th annual conference on Genetic and evolutionary computation
Year:
2012

Citing 7
Cited 5

Natural gradient works efficiently in learning

Neural Computation
Convergence results for the (1, λ)-SA-ES using the theory of ϕ-irreducible Markov chains

Theoretical Computer Science
Completely Derandomized Self-Adaptation in Evolution Strategies

Evolutionary Computation
Exponential natural evolution strategies

Proceedings of the 12th annual conference on Genetic and evolutionary computation
High dimensions and heavy tails for natural evolution strategies

Proceedings of the 13th annual conference on Genetic and evolutionary computation
General lower bounds for evolutionary algorithms

PPSN'06 Proceedings of the 9th international conference on Parallel Problem Solving from Nature
Performance analysis of the simultaneous perturbation stochastic approximation algorithm on the noisy sphere model

Theoretical Computer Science

Benchmarking separable natural evolution strategies on the noiseless and noisy black-box optimization testbeds

Proceedings of the 14th annual conference companion on Genetic and evolutionary computation
Benchmarking exponential natural evolution strategies on the noiseless and noisy black-box optimization testbeds

Proceedings of the 14th annual conference companion on Genetic and evolutionary computation
Investigating the impact of adaptation sampling in natural evolution strategies on black-box optimization testbeds

Proceedings of the 14th annual conference companion on Genetic and evolutionary computation
Benchmarking natural evolution strategies with adaptation sampling on the noiseless and noisy black-box optimization testbeds

Proceedings of the 14th annual conference companion on Genetic and evolutionary computation
Comparing natural evolution strategies to BIPOP-CMA-ES on noiseless and noisy black-box optimization testbeds

Proceedings of the 14th annual conference companion on Genetic and evolutionary computation

Quantified Score

Hi-index	0.00

Visualization

Abstract

This theoretical investigation gives the first proof of convergence for (radial) natural evolution strategies, on d-dimensional sphere functions, and establishes the conditions on hyper-parameters, as a function of d. For the limit case of large population sizes we show asymptotic linear convergence, and in the limit of small learning rates we give a full analytic characterization of the algorithm dynamics, decomposed into transient and asymptotic phases. Finally, we show why omitting the natural gradient is catastrophic.