Anzeige
Erschienen in:
31.03.2022
Learning to cooperate in agent-based control of queueing networks
Erschienen in: Queueing Systems | Ausgabe 3-4/2022
EinloggenAktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Excerpt
Control of queueing networks has been an active area of research for many years and with a variety of motivations, ranging from communication networks to supply chains. Nevertheless the problems amenable to clean and elegant analysis tend to be few and far between, and highly stylized. Real queueing networks tend to be plagued with many difficulties, both at the level of modeling and at the level of analysis, such as:
1.
Scale: One major issue is always the enormously large size of the network and the lack of structure because most of these networks are ‘emergent’ and structure, if any, can at best be characterized in statistical terms.
2.
Distributed, asynchronous control: The control, be it routing, admission or rate control, is at node or edge level, i.e., local and with local information. The only information available is the rewards and the local state. Gathering additional information through message passing may be unrealistic or possible only in a limited sense.
3.
Modeling issues: Assumptions about input processes and aspects of queueing dynamics, such as distributional assumptions or assumptions regarding independence or Markovianity, are often oversimplifications.
4.
Closed-loop stability: Despite major successes such as the backpressure scheme, one cannot say that the final word has been said on this.
5.
Choice of objectives: The primary objective of each queue is always its own throughput, but there can be secondary objectives such as energy efficiency and overall fairness. Also, ‘optimality’ is usually a tall order and one has to seek a ‘satisficing’ solution in the sense of Simon [13], i.e., one that meets some minimum specifications.
6.
Choice of policies: Distributed control of queueing networks is often viewed as a network game, thereby putting it firmly in the framework of stochastic games [11], with its equilibrium concepts such as Markov perfect equilibrium and Bayesian Nash equilibrium. However, it is not a priori ruled out that choice of a non-stationary policy at one or more queues may give a strictly better performance for all.