Signaling for Decentralized Routing in a Queueing Network

TL;DR

This paper studies a decentralized routing problem in a two-station queueing network, showing that optimal signaling strategies are threshold policies based on shared information, despite complex implicit communication.

Contribution

It demonstrates that optimal signaling in a decentralized queueing system can be characterized by explicit threshold policies depending on common information.

Findings

Optimal signaling strategies are threshold policies.

Threshold policies depend on common knowledge between controllers.

Explicitly determined threshold policies for minimizing costs.

Abstract

A discrete-time decentralized routing problem in a service system consisting of two service stations and two controllers is investigated. Each controller is affiliated with one station. Each station has an infinite size buffer. Exogenous customer arrivals at each station occur with rate $\lambda$. Service times at each station have rate $\mu$. At any time, a controller can route one of the customers waiting in its own station to the other station. Each controller knows perfectly the queue length in its own station and observes the exogenous arrivals to its own station as well as the arrivals of customers sent from the other station. At the beginning, each controller has a probability mass function (PMF) on the number of customers in the other station. These PMFs are common knowledge between the two controllers. At each time a holding cost is incurred at each station due to the customers waiting at that station. The objective is to determine routing policies for the two controllers that minimize either the total expected holding cost over a finite horizon or the average cost per unit time over an infinite horizon. In this problem there is implicit communication between the two controllers; whenever a controller decides to send or not to send a customer from its own station to the other station it communicates information about its queue length to the other station. This implicit communication through control actions is referred to as signaling in decentralized control. Signaling results in complex communication and decision problems. In spite of the complexity of signaling involved, it is shown that an optimal signaling strategy is described by a threshold policy which depends on the common information between the two controllers; this threshold policy is explicitly determined.