General Information
NameDr. David Levanony
DepartmentDepartment of Electrical and Computer Engineering
Emaillevanony@ee.bgu.ac.il
Academic RankSenior lecturer


Education

 

D.Sc. 1992, Department of Electrical Engineering, Technion - IIT, Haifa, Israel.

Dissertation: Recursive Metho ds for Identification in Continuous-Time Sto chastic

Processes.

Advisors: Prof. A. Shwartz and Prof. O. Zeitouni.

M.Sc. 1988, Department of Aeronautical Engineering, Technion - IIT, Haifa, Israel.

Thesis: Adaptive Parameter Estimation in Nonlinear Systems.

Advisor: Prof. N. Berman.

B.Sc. 1985, Cum Laude, D epartment of Aeronautical Engineering, Technion - IIT, Haifa,

Israel.

   

Research Interests

 Stochastic Adaptive Control;Adaptive Filtering;System Identification;Stability and Performance of Recursive Algorithms; Large Deviations Techniques and Applications.
   

Research Projects

 Adaptive performance optimization of stochastic systems in the presenceof parametric and structural uncertainties;Filtering and control problems.
   

Research Abstract

 Stochastic LQ Adaptive Control (with P. E. Caines, McGill University, Canada) A novel constrained optimization ap-proach is developed to overcomethe well known problem of suboptimal performance in standard LQ(linear quadratic) adap-tive control which results due to insufficient excitation. Opti-mal (long run) performance is obtained via a recursivesolution of a (time dependent) constrained optimization problem, pro-ducinga ML (maximum likelihood) type estimate which is biased towards lower control costs. Linear Adaptive Filtering Optimal adaptive (linear) filtering under system parametric uncertaintiesis sought-after.First, the geometric structure of ML limit parameter sets is character-ized.Then, based on the geometric study, a recursive adaptive filtering scheme is to be derivedto obtain (asymptotically) op-timal state estimates. An applicationstudy in GPS-aided Inertial Navigation is planned. Large Deviation Laws in Recursive Estimation This work is designed to extend partial LD (large deviations) results (see above), into complete LD laws for martingale LLNs. Those would provide a powerful tool for the evaluation of conver-gence rates of stochastic processes exhibiting a LLN type be-havior, most notably, some widely used recursive parameter estimation algorithms. Unlike earlier results which where ob-tained by using Gaussian methods, a derivation of complete LD laws would require a different approach based on the general LD theory. It is worth noting that, unlike standard LD results, conditional LD laws are sought-after here. Such conditional laws would enable to infer the performance to go from past observations. This in turn would lead to efficient stopping rules.