TU Darmstadt / ULB / TUprints

Modeling and Simulation of Multi-Task Networks Using Adaptation and Learning

Khawatmi, Sahar (2017)
Modeling and Simulation of Multi-Task Networks Using Adaptation and Learning.
Technische Universität Darmstadt
Ph.D. Thesis, Primary publication

PhD Dissertation - Text
Khawatmi.pdf - Accepted Version
Copyright Information: CC 1.0 Universal - Public Domain:the author relinquishes on all properities rights.

Download (8MB) | Preview
Item Type: Ph.D. Thesis
Type of entry: Primary publication
Title: Modeling and Simulation of Multi-Task Networks Using Adaptation and Learning
Language: English
Referees: Zoubir, Prof. Dr. Abdelhak M. ; Sayed, Prof. Dr. Ali H.
Date: 20 September 2017
Place of Publication: Darmstadt
Date of oral examination: 20 September 2017

This PhD thesis focuses on cooperative multi-task networks. Cooperative networks consist of a collection of agents with adaptation and learning abilities. The idea of sharing data among the neighboring agents is the basic tool for designing distributed algorithms for cooperative networks without a fusion center. This key technique is inspired by the collective behavior of some animal groups such as bee swarms, bacteria colonies, starling flocks, and fish schools. In these cases and in many more, the group of individuals as a whole exhibits a behavior that cannot be accessed at the individual members. This organized behavior can be understood by considering the large amount of interactions among agents.

There arises the need in many network applications to infer and track different models of interest in an environment. In our research, we focus on multi-task networks where the individual agents might be interested in different objectives. One challenge in these networks is that the agents do not know beforehand which models are being observed by their neighbors. Furthermore, the total number of the observed models and their indices are not available to them, either. We propose a distributed clustering technique that allows the agents to learn and form their clusters from streaming data in a robust manner. Once clusters have been formed, cooperation among agents with similar objectives can increase the performance of the inference task. Based on the cluster formation, the unused links among the agents that track different models are exploited to link the agents that are interested in the same model but do not have direct links between each other. We analyze the performance of the clustering scheme and show that the clustering error probabilities decay exponentially to zero. In addition, we examine the mean-square performance of the proposed clustering scheme. Furthermore, we propose a distributed labeling system, which ensures that each cluster has a unique index for its observed model.

Certain types of animal groups, such as bee swarms, consist of informed and uninformed agents where only the informed agents collect information about the environment. We consider a network where the informed agents observe different models and send information about them to the uninformed ones. Each uninformed agent responds to one informed agent and joins its group. We suggest an adaptive and distributed clustering and partitioning approach that allows the informed agents in the network to be clustered into different groups according to the observed models; then we apply a decentralized strategy to split the uninformed agents into groups of approximately equal size around the informed agents.

In some other situations, the agents in the network need to decide between multiple options, for example,to track only one of multiple food sources. We propose a distributed decision-making approach over adaptive networks where agents in the network collect data generated by different models. The agents need to decide which model to estimate and track.Once the network reaches an agreement on one desired model, the cooperation among the agents enhances the performance of the estimation task by relaying data throughout the network.

We investigate all scenarios and approaches in both cases: static and mobile networks.The simulations illustrate the performance of the proposed strategies and compare them with state-of-the-art approaches.

Alternative Abstract:
Alternative AbstractLanguage

Die vorliegende Doktorarbeit beschäftigt sich mit kooperativen Netzwerken mit mehreren Aufgaben. Kooperative Netzwerke bestehen aus einer Sammlung von Agenten mit Adaptions- und Lernfähigkeiten. Die Idee, Daten zwischen benachbarten Agenten auszutauschen, ist das grundlegende Mittel, um verteilte Algorithmen für kooperative Netzwerke ohne Fusionszentrum zu entwerfen. Dieses Schlüsselverfahren wurde durch das kollektive Verhalten diverser Tiergruppen, wie beispielsweise Bienenschwärme, Bakterienkolonien, Staren- und Fischschwärme, inspiriert. In diesen und vielen anderen Fällen weist die Gruppe als Ganzes ein Verhalten auf, das an den Individuen selbst nicht beobachtet werden kann. Diesem organisierten Verhalten kann auf den Grund gegangen werden, wenn man die Vielzahl an Interaktionen zwischen den Agenten betrachtet.

URN: urn:nbn:de:tuda-tuprints-68275
Classification DDC: 600 Technology, medicine, applied sciences > 620 Engineering and machine engineering
Divisions: 18 Department of Electrical Engineering and Information Technology > Institute for Telecommunications > Signal Processing
Date Deposited: 12 Oct 2017 08:29
Last Modified: 12 Oct 2017 08:29
URI: https://tuprints.ulb.tu-darmstadt.de/id/eprint/6827
PPN: 417732074
Actions (login required)
View Item View Item