- author = "Moshe Sipper and Weixuan Fu and Karuna Ahuja and Jason H. Moore",
- title = "From {MEGATON} to {RASCAL}: Surfing the Parameter Space of Evolutionary Algorithms",
- year = "2017",
- month = "13 " # jun,
- keywords = "genetic algorithms, genetic programming, Evolutionary Algorithms, Meta-Genetic Algorithm, Parameter Tuning, Hyper-Parameter",
-
URL = "
https://arxiv.org/abs/1706.04119",
- size = "17 pages",
- abstract = "The practice of evolutionary algorithms involves a mundane yet inescapable phase, namely, finding parameters that work well. How big should the population be? How many generations should the algorithm run? What is the (tournament selection) tournament size? What probabilities should one assign to crossover and mutation? All these nagging questions need good answers if one is to embrace success. Through an extensive series of experiments over multiple evolutionary algorithm implementations and problems we show that parameter space tends to be rife with viable parameters. We aver that this renders the life of the practitioner that much easier, and cap off our study with an advisory digest for the weary",
-
notes = "p14 ' -Don't spend too much time and resources on
tuning hyper-parameters.
- Random search is a good choice for said tuning.
- Robustness to hyper-parameter tuning is a desired quality of an evolutionary algorithm ...
- ... and if your algorithm requires very specific parameters, the chances of finding them are most likely slim at best. You're essentially in a needle-in-a-haystack situation in hyper-parameter space. Rethink and regroup.'
See blog entry: http://gpemjournal.blogspot.co.uk/2017/06/parameters-parameters-parameters_20.html
See \cite{Sipper2018}",
