Studies in Continuous Black-box Optimization

Studies in Continuous Black-box OptimizationThis PhD dissertation presents a collection of novel, state of the art algorithms for solving problems in the class of continuous black box optimization. Natural Evolution Strategies are a family of algorithms that constitutes a general purpose approach. Maintaining a parameterized distribution on the set of solution candidates, the natural gradient is used to update the distribution's parameters in the direction of higher expected fitness. Techniques

This PhD dissertation presents a collection of novel, state-of-the-art algorithms for solving problems in the class of continuous black-box optimization. Natural Evolution Strategies are a family of algorithms that constitutes a general-purpose approach. Maintaining a parameterized distribution on the set of solution candidates, the natural gradient is used to update the distribution's parameters in the direction of higher expected fitness. Techniques are introduced that addresses issues of convergence, robustness, computational complexity and algorithm speed. We also demonstrate how the principle of artificial curiosity can guide exploration in the context of costly optimization, introducing a response surface method that estimates the interestingness of each candidate point using Gaussian process regression. The results show best published performance on various standard benchmarks, as well as competitive performance on others.