Evolutionary Learning for Intelligent Automation: A Case Study

Abstract

Industrial automation calls for behavioral intelligence, that is, a mixture of flexibility,
\nrobustness and adaptiveness of robot behavior. We argue that efficient machine
\nlearning techniques can be a valuable tool for achieving behavioral intelligence. As a
\ncase study we apply ALECSYS, an implementation of a learning classifier system on a
\nnet of transputers, to a gross-motion problem for an industrial manipulator (an IBM
\n7547 with a SCARA geometry). A simple simulation environment allows us to
\nexperiment with different sensor configurations, and to obtain a first, coarse
\napproximation of the robot’s controller through learning. The controller is
\nsubsequently refined through a learning session run on the physical robot. As a
\nwhole, our work demonstrates some interesting distinctive features of the evolutionary
\ncomputation approach, viewed as a possible alternative to classical methods of
\nsoftware development.