Institute of Information Theory and Automation

You are here

Approximation of fully-probabilistic version of dynamic programming as a basis of universal learning, decision-making and control systems

Type of Work: 
dissertation
Affiliation/Phone: 
ÚTIA AV ČR, v.v.i., AS department, 266052274
Keywords: 
Adaptive systems, Bayesian learning and decision making, fully probabilistic design of decision strategies, approximation of mulrivariate implicit functions

Fully probabilstic design of dynamic decision startegies is a well-developed theoretical basis of learning decision systems, which are potentially widely applicable in technology, natural and societal systems. The applicability is strongly constrained by complexity of the associated optimisation, a special version of dynamic programming. In the considered case, it is necessary to approximate a scalar function of many variables, which is implictily described as a  solution of of non-linear integral-difference equation. The solution of this hard problem can be split into topics of several Phd theses, which will differ in the stress on functional analysis, approximation of functions or various heuristic methods encountered in connection with artificial intelligence. Also, software or simulation oriented solutions are welcome.   

Note: 
It can be solved at FJFI or FEL CVUT Prague, ZUC Plzen or elsewhere after agreement.
Bibliography: 
  1. M. Kárný, T.V.Guy, Fully probabilistic control design, Systems & Control Letters, 55:4, 259-265, 2006
  2. M. Kárný et al, Optimized Bayesian Dynamic Advising: Theory and Algorithms, Springer, London
  3. J. Si et al, Handbook of Learning and Approximate Dynamic Programming, Wiley-IEEE Press, Danvers, 2004, ISBN 0-471-66054-X
2018-08-13 09:51