Dagstuhl Seminar Proceedings, Volume 9371
Dagstuhl Seminar Proceedings
DagSemProc
https://www.dagstuhl.de/dagpub/1862-4405
https://dblp.org/db/series/dagstuhl
1862-4405
Schloss Dagstuhl – Leibniz-Zentrum für Informatik
9371
2010
https://drops.dagstuhl.de/entities/volume/DagSemProc-volume-9371
09371 Abstracts Collection – Algorithmic Methods for Distributed Cooperative Systems
From 06.09.09 to 11.09.09 the Dagstuhl Seminar 09371
``Algorithmic Methods for Distributed Cooperative Systems'' was held
in Schloss Dagstuhl~--~Leibniz Center for Informatics.
The purpose of this workshop was to bring together
researchers from different disciplines whose central
interest is in both algorithmic foundations and application
scenarios of distributed cooperative systems.
During the seminar, several participants presented their current
research, and ongoing work and open problems were discussed. Abstracts of
the presentations given during the seminar as well as abstracts of
seminar results and ideas are put together in this paper. The first section
describes the seminar topics and goals in general.
Links to extended abstracts or full papers are provided, if available.
Algorithms
cooperative systems
sensor networks
multi-robot systems
applications
1-16
Regular Paper
Sándor
Fekete
Sándor Fekete
Stefan
Fischer
Stefan Fischer
Martin
Riedmiller
Martin Riedmiller
Suri
Dubhash
Suri Dubhash
10.4230/DagSemProc.09371.1
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode
Cooperative Multi-Agent Systems from the Reinforcement Learning Perspective – Challenges, Algorithms, and an Application
Reinforcement Learning has established as a framework that
allows an autonomous agent for automatically acquiring – in a
trial and error-based manner – a behavior policy based on a
specification of the desired behavior of the system.
In a multi-agent system, however, the decentralization of the
control and observation of the system among independent agents
has a significant impact on learning and it complexity.
In this survey talk, we briefly review the foundations of
single-agent reinforcement learning, point to the merits and
challenges when applied in a multi-agent setting, and illustrate
its potential in the context of an application from the field
of manufacturing control and scheduling.
Multi-agent reinforcement learning
decentralized control
job-shop scheduling
1-5
Regular Paper
Thomas
Gabel
Thomas Gabel
10.4230/DagSemProc.09371.2
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode
Local Algorithms: Self-Stabilization on Speed
An introduction to distributed algorithms, in particular local algorithms. Essentially a practice talk of my SSS 2009 invited talk.
Local Algorithms
Self-Stabilization
Lower Bounds
Upper Bounds
MIS
1-18
Regular Paper
Christoph
Lenzen
Christoph Lenzen
Jukka
Suomela
Jukka Suomela
Roger
Wattenhofer
Roger Wattenhofer
10.4230/DagSemProc.09371.3
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode
On the Fairness of Probabilistic Schedulers for Population Protocols
We propose a novel, generic definition of emph{probabilistic schedulers} for population protocols. We design two new schedulers, the emph{State Scheduler} and the emph{Transition Function Scheduler}. Both possess the significant capability of being emph{protocol-aware}, i.e. they can assign transition probabilities based on information concerning the underlying protocol. We prove that the proposed schedulers, and also the emph{Random Scheduler} that was defined by Angluin et al. cite{AADFP04}, are all fair with probability $1$. We also define and study emph{equivalence} between schedulers w.r.t. emph{performance} and emph{correctness} and prove that there exist fair probabilistic schedulers that are not equivalent w.r.t. to performance and others that are not equivalent w.r.t. correctness. We implement our schedulers using a new tool for simulating population protocols and evaluate their performance from the viewpoint of experimental analysis and verification. We study three representative protocols to verify stability, and compare the experimental time to convergence with the known complexity bounds. We run our experiments from very small to extremely large populations (of up to $10^{8}$ agents). We get very promising results both of theoretical and practical interest.
Population Protocols
Fairness
Probabilistic Schedulers
Communicating Automata
Sensor Networks
Experimental Evaluation
1-23
Regular Paper
Ioannis
Chatzigiannakis
Ioannis Chatzigiannakis
Shlomi
Dolev
Shlomi Dolev
Sándor
Fekete
Sándor Fekete
Othon
Michail
Othon Michail
Paul
Spirakis
Paul Spirakis
10.4230/DagSemProc.09371.4
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode
On the Fundamental Limits of Broadcasting in Wireless Mobile Networks
In this talk, we investigate the fundamental properties of broadcasting in mobile wireless networks. In particular, we characterize broadcast capacity and latency of a mobile network, subject to the condition that the stationary node spatial distribution generated by the mobility model is uniform. We first study the intrinsic properties of broadcasting, and present a broadcasting scheme, called RippleCast, that simultaneously achieves asymptotically optimal broadcast capacity and latency, subject to a weak upper bound on the maximum node velocity. This study intendedly ignores the burden related to the selection of broadcast relay nodes within the mobile network, and shows that optimal broadcasting in mobile networks is, in principle, possible. We then investigate the broadcasting problem when the relay selection burden is taken into account, and present a combined distributed leader election and broadcasting scheme achieving a broadcast capacity and latency which is within a $Theta((log n)^{1+frac{2}{alpha}})$ factor from optimal, where $n$ is the number of mobile nodes and $alpha>2$ is the path loss exponent. However, this result holds only under the assumption that the upper bound on node velocity converges to zero (although with a very slow, poly-logarithmic rate) as $n$ grows to infinity.
To the best of our knowledge, our is the first paper investigating the effects of node mobility on the fundamental properties of broadcasting, and showing that, while optimal broadcasting in a mobile network is in principle possible, the coordination efforts related to the selection of broadcast relay nodes lead to sub-optimal broadcasting performance.
Wireless networks
mobile networks
broadcast capacity
broadcast latency
SINR interference model
1-9
Regular Paper
Giovanni
Resta
Giovanni Resta
Paolo
Santi
Paolo Santi
10.4230/DagSemProc.09371.5
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode
Stabilizing Consensus with the Power of Two Choices
Consensus problems occur in many contexts and have therefore been intensively studied in the past. In the standard consensus problem there are n processes with possibly different input values and the goal is to eventually reach a point at which all processes commit to exactly one of these values. We are studying a slight variant of the consensus problem called the stabilizing consensus problem. In this problem, we do not require that each process commits to a final value at some point, but that eventually they arrive at a common value without necessarily being aware of that. This should work irrespective of the states in which the processes are starting. Coming up with a self-stabilizing rule is easy without adversarial involvement, but we allow some T-bounded adversary to manipulate any T processes at any time. In this situation, a perfect consensus is impossible to reach, so we only require that there is a time point t and value v so that at any point after t, all but up to O(T) processes agree on v, which we call an almost stable consensus. As we will demonstrate, there is a surprisingly simple rule for the standard message passing model that just needs O(log n loglog n) time for any sqrt{n}-bounded adversary and just O(log n) time without adversarial involvement, with high probability, to reach an (almost) stable consensus from any initial state. A stable consensus is reached, with high probability, in the absence of adversarial involvement.
Distributed consensus
1-21
Regular Paper
Benjamin
Doerr
Benjamin Doerr
Leslie Ann
Goldberg
Leslie Ann Goldberg
Lorenz
Minder
Lorenz Minder
Thomas
Sauerwald
Thomas Sauerwald
Christian
Scheideler
Christian Scheideler
10.4230/DagSemProc.09371.6
Creative Commons Attribution 4.0 International license
https://creativecommons.org/licenses/by/4.0/legalcode