Combining Genetic Programming and Semi-quantitative Representation to Learn Models of Physical Systems

Scientific discovery implies exploring a vast search space of possible hypotheses in the hope of finding a model befitting the available data. Our present aim is to use Genetic Programming (GP) and a representation involving both crisp numbers and fuzzy quantity spaces to learn models of simple physical systems from a set of imperfect and incomplete data. We use the ECJ framework, to learn models of increasing complexity (u-tube, coupled tanks, and cascading tanks). The best fitness is obtained when a model covers all positive examples. Results show that the system can approximate the target models and that the use of a weighted fitness function seems to accelerate the learning process.