PUB - Publikationen an der Universität Bielefeld

We argue that endowing machines with perception and action raises many problems for which solutions derived from first principles are unfeasable or too costly and that artificial neural networks offer a worthwhile and natural approach towards a solution. As a result, learning methods for creating mappings or dynamical systems that implement a desired functionality replace the implementation of algorithms by programming. We discuss two neural network approaches, the Local Linear Maps (LLMs) and the Parametrized Self-Organizing Maps (PSOMs), that are well suited for the rapid construction of mapping modules and illustrate some of their possibilies with examples from robot vision and manipulator control. We also address some more general issues, such as the need for mechanisms of attention control and for the flexible association of continuous degrees of freedom, together with an ability for handling varying constraints during the selection of actions.