The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

Research Paper #925

Title:Evolving Complex Visual Behaviours using Generic Programming and Shaping
Authors:Perkins,SJ; Hayes,GM
Date:Aug 1998
Presented:Presented at the European Workshop on Learning Robots 1998, Edinburgh
Abstract:Learned or evolved robot controllers are often criticized for being unable to compete with hand-designed controllers. Hierarchical structuring of the controller coupled with staged learning can enhance the capabilities of learned controllers, but is subject to the criticism that the human designer is in effect doing all the hard work, and hence that this approach will not scale up. We suggest imposing an explicit barrier between the human designer and the internal workings of the controller, limiting the human to using external methods for influencing the learning process that might be more scalable. Such methods are similar to what is called "shaping" in animal training. In this paper we first introduce a simple but powerful GP-like learning architecture. We than enhance this by incorporating it initially into a rule-based system, and then into a multi-agent system. The utility of shaping techniques is then demonstrated with this architecture on a visual tracking task. robot controllers are evolved initially in simulation and then are transferred to a real robot. Performance comparisons are made with a hand-crafted controller for the same task.

[Search These Pages] [DAI Home Page] [Comment]