90 likes | 256 Views
Karel Macek, Department of Mathematics, FPENS, CTU Prague. osoom - Optimal Switching of optimizing methods. Ouline. Optimization and its issues Common sense practices OSOOM Algorithm Used submethods Results on test functions. Optimization and its issues. Optimization problem
E N D
Karel Macek, Department of Mathematics, FPENS, CTU Prague osoom -Optimal Switching of optimizing methods
Ouline • Optimization and its issues • Common sense practices • OSOOM Algorithm • Used submethods • Results on test functions
Optimization and its issues • Optimization problem • let S Rn and f:S→R • the aim is to find x0 S, so f(x0) is maximal • Issues • Function f • not differentiable (e.g. max) • not explicitely defined • Set S • shape (not continuous, not smooth) • dimensionality
Common sense practices • An optimizer (human operator) shlould • master more optimizing methods • have experience with them • remember the most successful ones • sometimes try to apply also not so succesful ones
OSOOM Algorithm • Formalization of above mentioned ideas • In more detail: • We have a population of individuals in Rn • A set of optimizing methods is available • In a step of optimization, all individuals use their an optimizing method with respect to a probability distribution, depending on • Individual • Previous method applied • Successfullity of this decision in previous time steps
Used submethods • Threshold Acceptance • Differential Evolution • Particle Swarm Optimization • Continuous Ant Colony Optimization • Conjugate Gradient
Results on test functions • Tested functions • Sphere (Dejonga 1) • Rosenbrock sattle (Dejonga 2) • Results • OSOOM is so good or better than single optimization methods
Conclusion • Discussion of typical issues related to optimization • Sketching the OSOOM algorithm • Presentation of obtained results