DocumentCode :
2251533
Title :
Two time-scale stochastic approximation for constrained stochastic optimization and constrained Markov decision problems
Author :
Tadic, Vladislav B. ; Doucet, Arnaud ; Singh, Sumeetpal
Author_Institution :
Dept. of Electr. & Electron. Eng., Melbourne Univ., Parkville, Vic., Australia
Volume :
6
fYear :
2003
fDate :
4-6 June 2003
Firstpage :
4736
Abstract :
In this paper, constrained stochastic optimization problems are considered for the case where the constraint functions are convex (but the criterion function can be non-convex) and the criterion and constraint functions are available only through their noisy observations. A general algorithm of the two time-scale stochastic approximation type is proposed for these problems. The proposed algorithm is applied to Markov decision problems with average cost, average constraints and parameterized stationary policy. The asymptotic behavior of the proposed algorithm is analyzed for the case where the algorithm step-sizes are constant and the noise in the observations of the criterion and constraint functions depends on the algorithm iterates.
Keywords :
Markov processes; approximation theory; decision theory; iterative methods; optimisation; algorithm iterates; asymptotic behavior; average constraints; average cost; constant algorithm step-sizes; constrained Markov decision problems; constrained stochastic optimization; convex constrained functions; noisy observations; nonconvex criterion function; parameterized stationary policy; two-time scale stochastic approximation; Approximation algorithms; Constraint optimization; Constraint theory; Convergence; Costs; Lagrangian functions; Optimization methods; Postal services; Stochastic processes; Stochastic resonance;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
American Control Conference, 2003. Proceedings of the 2003
ISSN :
0743-1619
Print_ISBN :
0-7803-7896-2
Type :
conf
DOI :
10.1109/ACC.2003.1242471
Filename :
1242471
Link To Document :
بازگشت