Family bootstrapping: A genetic transfer learning approach for onsetting the evolution for a set of related robotic tasks

Abstract

Studies on the bootstrap problem in evolutionary robotics help lifting the barrier from the way to evolve robots for complex tasks. It remains an open question, though, how to reduce the need for designer knowledge when devising a bootstrapping approach for any particular complex task. Transfer learning may help reducing this need and support the evolution of solutions to complex tasks, through task relatedness. Relying on the commonalities of similar tasks, we introduce a new concept of Family Bootstrapping (FB). FB refers to the creation of biased ancestors that are expected to onset the evolution of "a family" of solutions not just for one task, but for a set of related robot tasks. A general FB paradigm is outlined and the unique potential of the proposed concept is discussed. To highlight the validity of the FB concept, a simple demonstration case, concerning the evolution of neuro-controllers for a set of robot navigation tasks, is provided. The paper is concluded with some suggestions for future research.