Tuesday, July 20, 2010

To model or not to model?

How not to model autonomus behavior: "Models and instantiations are very different things. An instantiation is supposed to be a proper member of the class of phenomena under study. By contrast, a model need not be; it can be crude – almost ignoring the majority of interested aspects of the phenomenon of interest – and yet be extremely useful. In general, simple models tend to be scientifically very powerful. This is because the purpose of a model is not to replicate a phenomenon, but to help explain it. There are lots of ways in which this can happen that do not involve producing an instantiation: models can show us the mistake in our assumptions, they can be explanatory rich in the way they actually fail to capture the phenomenon of interest, they can act as proofs of concept, they can generate novel hypotheses, and generally they can help re-organize complex ideas by exercising and questioning our intuitions."
"Evolutionary robotics (ER) is still proving a useful and open-ended method for exploring this increasingly less constraining role of the designer that may be required to achieve strong artificial autonomy. ER hands in the task of filling in design specifications pertaining to mechanisms, morphology, structural and functional organization to an automatic process of artificial evolution (Harvey et al. 1997, Nolfi & Floreano, 2000). Thus, instead of designing a robot that must explore the environment but should go to the green light when the battery is down, one can attempt to design a robot that more generally must keep the battery up during its explorations, or more implicitly, a robot that explores indefinitely."

No comments: