Seminar on Computational Learning and Adaptation


 
Equation Discovery for Dynamic Systems Identification

Saso Dzeroski
Jozef Stefan Institute
Ljubljana, Slovenia
www-ai.ijs.si/SasoDzeroski/

Machine discovery is concerned with the task of finding laws from experimental and/or observational data. Early machine discovery systems reconstructed scientific laws, such as the ideal gas law, formulated as algebraic equations. Such laws are typically found in high-school physics textbooks. In Ljubljana, we have developed a number of systems that discover differential equations describing the observed behavior(s) of a dynamic system. We have thus moved the application domain of machine discovery systems to university-level physics. This talk will give a brief overview of these developments. A number of issues have been addressed in the systems mentioned, such as handling noise and using existing domain knowledge to constrain the space of possible models. The search methods used include exhaustive search, greedy and beam search, and even genetic programming. Several practical problems have been addressed using the developed systems, including two ecological modelling problems. A recent development extends our approach to the discovery of partial differential equations, which are especially appropriate for modelling biological systems.


Date: Thurs., Sept 28

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule