Abstract
Model learning often implies exploring a vast search
space of possible hypotheses in the hope of finding a
solution. Qualitative model learners are mostly based
on Inductive Logic Programming (ILP), which is a
systematic method which tends to be well fitted for
exploring solutions in a narrow search space. We
present a semi-quantitative model learner that uses
Genetic Programming (GP), which is well suited for
exploring a broad search space. We learn simple
physical systems based on a formalism involving both
crisp numbers and fuzzy quantity spaces. We use the ECJ
framework,1 and the fitness of a model is set to be
optimal when it covers all positive examples. Several
experiments are performed to learn and reuse models of
physical systems of increasing complexity; firstly a
u-tube, then coupled tanks, and finally cascading
tanks. Results show that the system can approximate the
target models in reasonably good conditions, and that
there is still scope for optimisation.
Users
Please
log in to take part in the discussion (add own reviews or comments).