Sunday 4:30 PM–5:15 PM in Room 1

Evolutionary Algorithms: Perfecting the Art of "Good Enough"

Liz Sander

Audience level:
Intermediate

Description

Evolutionary algorithms let us tackle all kinds of impossible problems. Want to design a short delivery route, but there are more possible solutions than atoms in the universe? Well, evolutionary algorithms can't promise to find the optimal solution, but can guarantee finding a pretty great one. I'll give an overview of these algorithms, and how you can use them for your own impossible problems.

Abstract

Science and data analytics are full of optimization problems: we may want to minimize cost, maximize model fit, or find the best clustering for a dataset. Many problems are combinatorially explosive, making it hopeless to find the global optimum with an exhaustive search. But many algorithms can find excellent solutions. As long as you can measure the "fitness" of a given solution, you can use evolution to find increasingly better solutions. This talk will cover:

  • what an evolutionary algorithm is, and how to use them to solve scientific problems
  • fitness landscapes, mutation functions, and fitness functions
  • an overview of some useful heuristic optimizers (hill climbers, simulated annealing, MCMC, genetic algorithms), and their strengths and weaknesses

Sponsors

Become a sponsor.