Tracking and Assigning Jobs to a Markov Machine

TL;DR

This paper models a communication system with a machine and server, optimizing sampling and job assignment policies to minimize costs related to machine state accuracy and job dropping, using Markov decision processes.

Contribution

It introduces a Markov decision process framework for optimal sampling and job assignment in a system with state estimation and penalties, revealing threshold policies as optimal.

Findings

Optimal threshold policy identified for the system.

Necessary and sufficient conditions for policy optimality derived.

Optimal threshold computed without state space bounds.

Abstract

We consider a time-slotted communication system with a machine, a cloud server, and a sampler. Job requests from the users are queued on the server to be completed by the machine. The machine has two states, namely, a busy state and a free state. The server can assign a job to the machine in a first-in-first-served manner. If the machine is free, it completes the job request from the server; otherwise, it drops the request. Upon dropping a job request, the server is penalized. When the machine is in the free state, the machine can get into the busy state with an internal job. When the server does not assign a job request to the machine, the state of the machine evolves as a symmetric Markov chain. If the machine successfully accepts the job request from the server, the state of the machine goes to the busy state and follows a different dynamics compared to the dynamics when the machine goes to the busy state due to an internal job. The sampler samples the state of the machine and sends it to the server via an error-free channel. Thus, the server can estimate the state of the machine, upon receiving an update from the source. If the machine is in the free state but the estimated state at the server is busy, the sampler pays a cost. We incorporate the concept of the age of incorrect information to model the cost of the sampler. We aim to find an optimal sampling policy such that the cost of the sampler plus the penalty on the machine gets minimized. We formulate this problem in a Markov decision process framework and find how an optimal policy changes with several associated parameters. We show that a threshold policy is optimal for this problem. We show a necessary and sufficient condition for a threshold policy to be optimal. Finally, we find the optimal threshold without bounding the state space.