Evolving continuous optimisers from scratch

This work uses genetic programming to explore the space of continuous optimisers, with goal discovering novel ways doing optimisation. In order keep search broad, optimisers are evolved from scratch using Push, a Turing-complete, general-purpose, language. The resulting found be diverse, and their optimisation landscapes variety interesting, sometimes unusual, strategies. Significantly, when applied problems that were not seen during training, many generalise well, often outperform existing optimisers. supports idea effective forms can discovered in an automated manner. paper also shows pools hybridised further increase generality, leading perform robustly over broad problem types sizes.