Which Dynamic Systems?

These systems are composed by interacting agents (devices and/or software elements) that are connected through a network, within an environment subject to constraints, policies or potentially changing operating conditions. Agents are uncoordinated or loosely coordinated, and have a partial, imperfect (error- or loss-prone) visibility of their networking environment; decisions are based on this partial, imperfect visibility.

Examples of these systems include data transfer and the provision of distributed services (in data centers, opportunistic deployments, or distribution or access networks), the establishment of reliable connectivity (transit networks and infrastructures), or the dynamic distribution of computing resources (e.g., IOT deployments in the edge).

Some Use Cases

Reliable multicast, that is, reliable content distribution from one source to a set of destinations, connected through multi-hop networks with variable topology and dynamic operating conditions (node capacities, available speeds, caching policies). The distributed reliability of multicast flows in this context depends on the ability of destination nodes to (i) detect losses, and (ii) identify relay nodes and destinations capable and ready to repair detected losses.  In these contexts, which naturally arise in “rescue” scenarios and/or ad-hoc emergency deployments to respond to disasters or natural disasters, the establishment of “command and control” services can require that the distributed agents establish by themselves, or with a weak coordination from the partial and sometimes imprecise data at their disposal, satisfactory balances with regard to the information flow from source(s) to destinations, on one hand, and adaptive reliability mechanisms in the event of losses, on the other. The overall responsiveness of these systems to changes in the environment (in network topology or in the capabilities or demands of the devices involved), the speed of implementing a concrete forwarding or “caching” policy (for privilege or avoid certain regions or devices of the network), and also the robustness of collective decisions in the face of occasional or transient disturbances of the data taken into consideration by the agents, are crucial elements of the dynamic performance to be studied and optimized.
Discovery and dynamic placement of resources. It is possible to envisage multi-agent systems in networks, with agents subject to weakly centralized data synchronization needs (transmitted from a central entity to a subset of network devices equipped with storage and “forwarding” capacities , which would make it happen to all the rest); varying demands on compute resources, for which optimal placements may change over time (depending on observed field conditions); or established capacities or fixed placement policies to which consumer devices should automatically adapt. This scenario could be relevant for opportunistic deployments, extensive IOT networks, or communication networks deployed on large infrastructures subject to variable demands and constraints (railways, electrical network).

Objectives

The objective of this project is to understand the emerging behaviors of these dynamic systems, and evaluate and improve the performance of simple, decentralized, lightweight and easy-to-implement architectures. This requires the analysis and optimization of the performance, stability and sensitivity of agent learning and decision-making mechanisms involved on this type of system. We focus on aspects related to the responsiveness of these systems to changes in operating conditions and/or agent policies: change detection, cost of phase transitions related to these changes, time before the re-establishment of new system equilibria, and quality of stationary configurations.

DYNSYS focuses on the study of the reactivity of these networked systems in the face of breakdowns, policy changes or variations in the constraints (demands, capacities) and operating conditions of the agents, and the robustness stationary configurations in the face of transient or occasional degradations of these conditions. Other aspects of interest include the dynamic characteristics of these systems, including convergence and the existence of equilibria dependent on external conditions, and the quality of the decisions taken during transient states.