 |
Center for Nonlinear Studies |
Coevolutionary learning is a biologically inspired machine learning technique in which a population of models or learners coevolves with a population of data or training examples. The biological inspiration comes from host-parasite (or predator-prey) competitions, in which species A evolves to thwart attacks from species B, only to see species B evolve to outwit the adaptations of species A, and so on, in an ever spiraling "biological arms race". Likewise (the hope goes), in coevolutionary learning, learners will evolve to successfully classify the current population of training examples, but in turn the training examples will evolve to be increasingly challenging for the learners, who in turn will become even more sophisticated, and so on. The (hoped for) result will be more successful learning with a smaller number of training examples than in traditional techniques. Coevolutionary learning first achieved popularity in the 1990s with the success of Danny Hillis's coevolving sorting networks. While the idea is compelling and while there have been other successful applications of coevolution on toy problems, there have also been numerous failures of the method, and there is still no good understanding of the conditions under which such methods will work well, and whether such methods can scale to more complicated problems. In this talk I will review the highlights of work on coevolution---both successes and failures---and present new research aimed at understanding the role of spatial distribution of populations in successful applications of coevolutionary learning.
Host: Christof Teuscher, CS-1