Simultaneous Perturbation Stochastic Approximation in Decentralized Load Balancing Problem∗∗The authors acknowledge the Russian Ministry of Education and Science (agreement 14.604.21.0035, unique no. RFMEFI60414X0035), RFBR (projects 13-07-00250, 14-08-01015, and 15-08-02640), and SPbSU (project 6.37.181.2014).
Natalia Amelina, Victoria Erofeeva, Oleg Nikolaevich Granichin, Мальковский Николай Владимирович · IFAC-PapersOnLine · 2015
In this work the load balancing problem is studied for decentralized stochastic network with unknown but bounded noise in measurements and varying productivities of agents. The load balancing problem is formulated as a consensus problem in a stochastic network. Consideration of Laplasian potential function corresponded to the network graph allows to introduce a new randomized local voting protocol with constant step-size which is based on simultaneous perturbation stochastic approximation algorithm. The conditions are formulated for the approximate consensus achievement which corresponds to achieving of a suboptimal level of agents' load. The new algorithm is illustrated by simulations.