Research Seminar

Spring 2020

Visualizing Dynamic Clustered Data Using Area-proportional Maps

  • Date: Fri, 3 Jul 2020, 13:00
  • Location: SR 301/Online
  • Speaker: Christian Schnorr, ITI Wagner; Master Thesis
  • Inhalt: We explore how cluster information in graphs can be visualized naturally as countries on a map in a dynamic setting where the graph changes over time. We want there to be a strong correlation between a country's size and the size of the cluster it represents, and want to be able to react to dynamic changes of the graph in a way that preserves the viewer's mental model of the map. This is a challenging problem because popular algorithms for visualizing graphs as geographic-like maps struggle with clusters being fragmented across different countries on the map, and because redrawing the map for a new, albeit similar, input graph does generally not preserve the viewer's mental map. To address these issues, we propose a framework that guarantees to keep clusters as continuous regions in the generated maps and supports dynamic inputs by allowing for small, incremental updates such as inserting or removing regions over time. We do this by working on an intermediate plane graph whose vertices correspond to clusters in the input graph, the cluster graph, and constructing the map as a contact representation of this cluster graph while accounting for the dynamic changes in both the cluster graph and its contact representation. This framework can be applied to a variety of real-world applications, allowing us to visualize clusters in the underlying data and how they change over time, thereby enabling viewers of the visualization to detect trends in the data easily.

Multilevel Hypergraph Partitioning with Vertex Weights Revisited

  • Date: Mon, 22 Jun 2020, 11:00
  • Speaker: Nikolai Maas, ITI Sanders; Bachelor Thesis
  • Inhalt: We analyze the k-way partitioning problem for hypergraphs with vertex weights and specifically the task of finding a balanced partition within the multilevel context. Problems from practical applications such as VLSI design and sparse matrix-vector multiplication naturally incorporate weighted instances, but the topic is only partially covered by current state-of-the-art partitioners. We investigate existing definitions of the balance constraint for weighted hypergraphs and propose a new generalized definition. In order to construct a partitioning algorithm that guarantees a balanced solution, we examine different aspects of the multilevel paradigm. In particular, additional measures are necessary for k-way partitioning via recursive bisection. To address this, we develop an approach based on an assignment of heavy vertices as fixed vertices to both blocks of the bisection and use theoretical results to prove that the balance of the final partition can be guaranteed. Our algorithm is integrated into the hypergraph partitioner KaHyPar and is evaluated in extensive computational experiments. Unlike the competing algorithms, our configuration always computes a balanced partition on the tested instances. The solution quality is equivalent to the current version of KaHyPar while slightly improving the running time.

Scalable Radiation Model Implementation Using Shortest-Path Distances

  • Date: Fri, 5 Jun 2020, 14:00
  • Speaker: Adrian Cierpka, ITI Wagner; Bachelor Thesis
  • Inhalt: Das Verkehrsaufkommen in einer Region zu bestimmen ist keine einfache Aufgabe. Es benötigt viele Daten und die Erfahrung von Experten um hochqualitative Verkehrsdaten zu berechnen. Somit ist die Erstellung solcher Daten teuer und zeitaufwändig. Meistens werden die Ergebnisse nicht veröffenlicht und so gibt es Bedarf für Algorithmen die ein ungefähres, realistisches Verkehrsaufkommen in einer Region berechnen. Dabei dürfen solche Algorithmen nur Daten benötigen, die öffentlich verfügbar sind und müssen einfach zu bedienen sein. Die Verkehrsdaten die man mit solchen Algorithmen berechnen kann, werden in Benchmarks verwendet um Verkehrsalgorithmen zu testen. Buchhold et al. haben in ihrer Arbeit Implementierungen des Radiation-Modells vorgestellt und mit diesen gute Ergebnisse erzielt. Zwei dieser Algorithmen heißen DRAD und TRAD. DRAD überzeugt mit guten Ergebnissen, ist aber langsamer als TRAD, welcher jedoch schlechtere Ergebnisse erzielt. In dieser Arbeit stellen wir einen Algorithmus vor, der wie TRAD eine Baum-basierte Suche verwendet. Anstatt georaphischer Distanzen nutzen wir jedoch kürzeste Pfad Distanzen. Dieser Algorithmus, CRAD genannt, liefert die gleiche Ergebinsqualität wie DRAD, ist aber in den meisten Experimenten deutlich schneller. Des Weiteren zeigen wir am Ende dieser Arbeit, dass wir nur mit Daten, die öffentlich verfügbar sind gute Ergebnisse erzielen können.

Min-cost flow algorithms for the Wind Farm Cabling Problem

  • Date: Fri, 8 May 2020, 14:00
  • Speaker: Marc Jenne, ITI Wagner; Bachelor Thesis
  • Inhalt: The Wind Farm Cabling Problem (WCP) describes the problem of finding a suitable cabling between a set of turbines and substations in wind farms with the goal of transmitting the whole electricity produced by the turbines to the substations. While there exist numerous possibilities of feasible cablings, it is of interest to find the one that minimizes the total costs of the required cables. This problem can be modeled as a Minimum-Cost Flow Problem (MCF), a well-known problem for which multiple algorithms already exist. These algorithms are proven to find an optimal solution in polynomial time; however, they cannot be used for the WCP without adaptions because of the primary difference between both problems: while the cost function is linear in the MCF, it is a non-linear step function in the WCP. We take a closer look at some of those algorithms and examine whether they can be adapted so that they are able to provide solutions for the WCP. We describe the algorithms and their different approaches to solve the MCF and outline the problems that arise once the cost function becomes non-linear. We present our adaption of the Successive Shortest Path Algorithm and provide multiple strategies to solve the WCP with this algorithm. In an experimental evaluation, we compare the best of these strategies to a NCC algorithm and an exact MILP solver in terms of running times and quality of the found solutions.

A Comparative Analysis of Switchings in Static and Dynamic Power Grids

  • Date: Tue, 28 Apr 2020, 13:00
  • Speaker: Adrian Grupp, ITI Wagner - Master's Thesis
  • Inhalt: The power grid is the largest machine humankind has ever built. However, this machine is facing wide-ranging structural change. The transition towards intensive usage of renewable energy sources introduces new challenges: How do we cope with their volatility and how can we use them as efficiently as possible? Exploiting structural properties may be one answer to this. On one hand it is known that purposefully removing transmission lines from a power grid, so called switching, can in fact resolve congestions and overloads, which can eventually lead to a higher power throughput. In the language of graph theory this can be formalized as an optimization problem on flow networks, the Maximum Transmission Switching Flow Problem. On the other hand the concern is that the power grid can be compromised by such big topological interventions as the removal of a whole transmission line. Stable operation has to be guaranteed at all times which means adhering to the network's utility frequency. The oscillator model, used in statistical physics, can be utilized to study the transient behavior of the power grid. With it we can investigate whether switchings still allow network synchronization. Additionally, small signal analysis can be used to analyze the stability of the system after the switching took place. In this work we give a comparative analysis of the two mentioned approaches. We study how optimization results from the static world of graph theory behave in the dynamic setting of real world power grids and assess if they are reasonable with regards to the stability of the network frequency. We want to shed light on the derivation of the models and give a first answer to the question whether we can safely use switching in order to improve our power grids. Where? DFNconf Meeting Room Name: Adrian Grupp Abschlussvortrag DFNconf Meeting Room Number: 979145276 DFNconf Meeting Link: https://conf.dfn.de/webapp/conference/979145276

Fall 2019

Acyclic n-Level Hypergraph Partitioning

  • Date: Thu, 19 Mar 2020, 10:00
  • Location: -
  • Speaker: Daniel Seemaier, ITI Sanders; Master Thesis
  • Inhalt: Directed acyclic graphs are widely used to model the data flow and execution dependencies of streaming applications. Efficient parallelization of such graphs requires acyclic partitioning of the dependency graph. However, normal graphs are not always capable of precisely modelling such dependencies. Thus, we consider directed acyclic hypergraphs (DAHs). In this work, we present the first n-level hypergraph partitioning algorithm for directed acyclic hypergraphs. Moreover, we show (i) that our algorithm beats the current state of the art for directed acyclic graph partitioning in terms of solution quality on realistic instances, (ii) that n-level algorithms are superior to single level algorithms, and (iii) that our algorithm improves on the makespan of a parallelized image streaming application.

PASAR – Planning as Satisfiability with Abstraction Refinement

  • Date: Wed, 5 Feb 2020, 14:00
  • Location: R -119
  • Speaker: Nils Froleyks, ITI Sanders; Master Thesis
  • Inhalt: We present a new algorithm for the satisficing (non-optimal) classical planning problem. The presented planner participated in the Sparkle Planning Challenge 2019 and won a considerable share (22.13%) of the first prize. We combine SAT-based planning with forward state space search using the principles of counterexample-guided abstraction refinement (CEGAR). A state-of-the-art SAT solver is used to solve an abstraction of the planning task at hand. As an abstraction, we use an incomplete encoding where interference between actions is ignored. With a graph-theoretic test we determine whether the solution found by the SAT solver can be directly transformed into a plan. If it can be transformed, we realize an encoding allowing the application of many actions in parallel (exist-step semantics) in a very compact manner, without imposing a fixed order on the actions. If the transformation is not possible, we use the solution to define a heuristic function that is used during a state space search. If the search fails to find a plan within a very short timeout, the abstraction is refined. To increase the synergy between SAT solving and forward search, both can learn new actions and add them to the problem. This allows the SAT solver to make big jumps in the search space, while the forward search benefits from the SAT solver’s ability to solve combinatorially hard problems. Using benchmark domains from recent international planning competitions, we compare our approach with various state-of-the-art planners from the fields of SAT-based planning and heuristic search. On almost all tested domains we can match the performance of the best tested solvers and on some domains we outperform the entire competition.

Improving the Compact Bit-Sliced Signature Index COBS for Large Scale Genomic Data

  • Date: Fri, 10 Jan 2020, 14:00
  • Location: SR 301
  • Speaker: Daniel Ferizovic, ITI Sanders; Master Thesis
  • Inhalt: In this thesis we investigate the potential for improving the Compact Bit-Sliced Signature Index (COBS) [BBGI19] for large scale genomic data. COBS was developed by Bingmann et al. and is an inverted text index based on bloom filters. It can be used to index k-mers of DNA samples or q-grams of plain text data and is queried using approximate pattern matching based on the k-mer (or q-gram) profile of a query. In their work Bingmann et al. demonstrated a couple of advantages COBS has over other state of the art approximate k-mer bases indices, some of which are extraordinary fast query and construction times, but as well as the fact that COBS can be constructed and queried even if the index does not fit into main memory. This is one of the reasons we decided to look more closely at some areas we could improve COBS. Our main goal is to make COBS more scalable. Scalability is a very important factor when it comes to handling DNA related data. This is because the amount of sequenced data stored in publicly available archives nearly doubles every year, making it difficult to handle even from the perspective of resources alone. We focus on two main areas in which we try to improve COBS. Those are index compression through clustering and distribution. The thesis presents our findings and improvements achieved in respect to those areas.

Weighted F-Free Edge Editing

  • Date: Fri, 13 Dec 2019, 14:00
  • Location: SR 301
  • Speaker: Jonas Spinner, ITI Wagner; Bachelor Thesis
  • Inhalt: The Weighted F-free Edge Editing Problem asks for a set of edge edits (insertions/deletions) with minimum cost so that the edited graph does not contain induced subgraphs from a given finite set F. We focus on adapting an exact FPT algorithm for unweighted editing to the weighted case and compare it against an ILP. We evaluate the algorithms and a set of speed-up techniques on protein-protein interaction networks for F = {C4, P4}. We compare the effect of lower bound algorithms, subgraph selection rules for branching and parameter search strategies on the FPT algorithm.

Solving geometric optimization problems by reinforcement learning: crossing minimization

  • Date: Wed, 11 Dec 2019, 13:00
  • Location: SR 301
  • Speaker: Yani Kolev, ITI Wagner; Master Thesis
  • Inhalt: Visualising relations between data points is an increasingly important problem in both everyday life and and numerous fields of scientific research. As the number of crossings is among the most important qualities of a graph drawing, over the years many approaches to minimising it have been developed. Due to the recent success of techniques based on Reinforcement Learning in the realm of game playing AI, we propose a new approach to crossing minimisation based on this technology. We base our algorithm on the Monte Carlo Tree Search, which utilises a Neural Network to guide the placement of each vertex in the graph drawing. Our framework is designed to allow easy adaptation to other similar domains. We train the network using a graph embedding tool - node2vec. Our results suggest that Monte Carlo Tree Search does indeed offer a large improvement on the crossing number over random chance, but the neural network we use to iteratively enhance it does not capitalise on the information provided by node2vec. We relive that incorporating concrete geometric data in the node2vec representation would vastly improve our results, as would alternate, more suitable embedding algorithms.

Routenplanung mit temporären Straßensperrungen und ortsabhängigen Wartekosten

  • Date: Fri, 29 Nov 2019, 14:00
  • Location: SR 301
  • Speaker: Jakob Wagenblatt, ITI Wagner; Bachelor Thesis
  • Inhalt: In dieser Arbeit untersuchen wir das Problem der Routenplanung auf einem Straßennetzwerk mit temporären Sperrungen. In vielen Ländern ist es Lastkraftwagen-Fahrern verboten, bestimmte Straßen in einem spezifizierten Zeitraum zu befahren. Zusätzlich existieren regional beschränkte Sperrungen, wie zum Beispiel Nachtfahrverbote in Städten. Aufgrund dieser Einschränkungen können Fahrer sich gezwungen sehen, auf Parkplätzen zu warten. Wir definieren in dieser Arbeit ein Modell, welches das Finden von geeigneten Parkplätzen bereits in der Routenplanung berücksichtigt. Außerdem priorisieren wir das Warten an Parkplätzen guter Qualität, um den Reisekomfort des Fahrers zu verbessern. Weiterhin minimieren wir sowohl die Ankunftszeit als auch die Fahrzeit. Da die beiden letztgenannten Kriterien jedoch nicht immer vereinbar sind, bieten wir gegebenenfalls mehrere mögliche Routen an. Wir entwerfen einen Algorithmus, der diese Routen effizient berechnet. Anschließend analysieren wir die Eigenschaften und Komplexität des Modells wie auch des Algorithmus. Die Laufzeit des Algorithmus verbessern wir mithilfe einer Kombination aus bewährten und neuen Beschleunigungstechniken. Dies führt zu einer durchschnittlichen Laufzeit von unter einer Sekunde für europaweite Anfragen.

Cube&Conquer-inspired Techniques for Parallel Automated Planning

  • Date: Fri, 8 Nov 2019, 14:00
  • Location: SR 301
  • Speaker: Jean-Pierre van der Heydt, ITI Sanders; Bachelor Thesis
  • Inhalt: The field of automated planning is concerned with the automatic generation of plans for different domains of problems. To allow the usage in as many fields as possible, only minimal assumptions are made to the domains. Since the complexity of these problems require a great amount of computation resources, the usage of multi core systems becomes more important. Nevertheless, there exists no clear best strategy to solve the planning problem in parallel. Cube and Conquer has proven to be a new and successful approach for solving problems of propositional logic in parallel. In this thesis we will translate the ideas of Cube and Conquer to automated planning. We focus on the realisation of cubes through nodes in the state space graph of the planning problem. In doing so, we present techniques to generate the cubes and assign computation time to them. Thereby we hope to approach parallel planning from a new angle. We are able to solve more problems than a comparable sequential algorithm. On hard test cases our algorithm scales well for up to eight cores. On easy test cases or when using more than eight cores we achieve no significant speedup.

Edge Guarding Plane Graphs

  • Date: Fri, 25 Oct 2019, 14:00
  • Location: SR 301
  • Speaker: Paul Jungeblut, ITI Wagner; Master's Thesis
  • Inhalt: Let G=(V,E) be a plane graph. We say that a face f is guarded by an edge vw if at least one of {v,w} is on the boundary of f. This talk presents the results of my master thesis giving bounds on the minimum number of edges needed to guard all faces of any n-vertex plane graph in some specific graph classes. After a general overview of previous and new results we will consider stacked triangulations in more detail: For the lower bound we construct particularly hard instances, for the upper bound we see how to locally reduce the size of a stacked triangulation to use induction on the number of vertices.

Improving Distributed External Sorting for Big Data in Thrill

  • Date: Fri, 11 Oct 2019, 14:00
  • Location: Raum 211
  • Speaker: Jonas Dann, ITI Sanders; Master Thesis
  • Inhalt: Big data analytics and scientific computing require processing ever growing amounts of data. Competition between companies and between academics applies pressure to analyze data ever faster. In recent years, big data processing frameworks like Apache Spark, Apache Flink and Thrill came into popularity to aid this development. Such frameworks simplify the algorithm development significantly, by hiding most of the complexities of working with distributed computer clusters and disks. This talk focuses on Thrill. One of the most important nontrivial algorithmic primitive in these frameworks is sorting. Sorting accounts for a significant percentage of computer use overall. This talk will present the state of the art distributed external sorting algorithm Canonical Merge Sort. Canonical Merge Sort however shows weaknesses when sorting streams of data, since it only uses local knowledge about the data set distribution when redistributing data. Thus, an improvement of the algorithm is proposed and implemented in Thrill. Experiments along several scale factors and with several data sets are conducted and compared with the original Canonical Merge Sort algorithm. Results show similar overall performance and improvements in network communication on certain data sets. Finally, next steps for the newly developed algorithm are discussed.

Simultaneous Integer Flows

  • Date: Fri, 4 Oct 2019, 14:30
  • Location: SR 301
  • Speaker: Emil Dohse, ITI Wagner; Bachelor's Thesis
  • Inhalt: Flow algorithms have important practical and theoretical applications. A recent trend in algorithmics is to look at simultaneous variants of well-known problems. In this thesis, simultaneous integer flows are examined. For a given family of graphs, some edges in common, and some individual ones, a simultaneous flow consists of a feasible flow for each graph of the family so that the flow is the same on all common edges. We start by comparing simultaneous flows to standard flows, to see that most structural properties, like augmenting paths or min-cut max-flow duality, do not seem to translate to simultaneous flows. We then consider the decisions problem SimFlow, that asks whether a simultaneous integer flow with a certain value exists. We show that SimFlow is NP-complete even with heavy restrictions to the flow networks, like the common network only consisting of a path. In addition to that, we show that there is no polynomial-time approximation algorithm for simultaneous flows with subexponential approximation factor. On the positive side, we prove that SimFlow is FPT in the number of common or individual edges.

Upward-Rightward-Prescribed Planarity

  • Date: Fri, 4 Oct 2019, 14:00
  • Location: SR 301
  • Speaker: Vera Chekan, ITI Wagner; Bachelor's Thesis
  • Inhalt: We introduce and investigate a new planarity variant for directed graphs called upward-rightward-prescribed planarity. An upward-rightward-prescribed graph is a directed graph in which every edge is assigned either a u- or an r-label. Such a graph is called upward-rightward-prescribed planar if there exists a planar drawing in which each u-labeled edge is drawn y-monotonic and each r-labeled edge is drawn x-monotonic. This planarity variant is strongly related to upward planarity and windrose planarity. We show that testing graphs for upward-rightward-prescribed planarity is NP-hard even for single-source graphs and consider a restricted setting where the subgraph induced by the u-labeled edges is a biconnected spanning subgraph and the embedding of this subgraph is fixed. For this restriction, we provide a linear-time algorithm that decides if an upward-rightward-prescribed drawing exists and in the positive case, constructs such a drawing. This algorithm reduces upward-rightward-prescribed-planarity testing to windrose-planarity testing. We show that there exist upward-rightward-prescribed planar graphs that admit no straight-line drawing. The relation to windrose planarity implies that every upward-rightward-prescribed planar graph has an (upward-rightward-prescribed planar) drawing on a grid of polynomial size with at most three bends per edge.

Spring 2019

Spaces of phylogenetic networks

  • Date: Fri, 27 Sep 2019, 14:30
  • Location: SR 301
  • Speaker: Jonathan Klawitter, ITI Wagner; Guest from the University of Auckland
  • Inhalt: Phylogenetic networks are leaf-labelled graphs used to visualise evolutionary histories of species. Using rearrangement operations that make small graph-theoretical changes to transform one network into another one, we obtain a graph over a set of phylogenetic networks. In this talk I will present the results of my PhD work on such spaces of phylogenetic networks, including results on connectedness, the size of neighbourhoods, and shortest paths.

Local Ordered Covering Numbers

  • Date: Fri, 27 Sep 2019, 14:00
  • Location: SR 301
  • Speaker: Laura Merker, ITI Wagner
  • Inhalt: Queue Number und Stack Number sind zwei Graphparameter, die in der Literatur viel untersucht wurden. Bei beiden Parametern wird ein Layout gesucht, das aus einer Knotenordnung und einer Partition der Kanten besteht, wobei bestimmte geordnete Subgraphen in keiner Teilmenge der Partition auftreten dürfen. Während klassischerweise die Größe der Partition minimiert wird, definieren wir eine lokale Variante der Parameter. Hierbei heißt ein Layout k-lokal, wenn jeder Knoten in höchstens k verschiedenen Teilmengen enthalten ist. Gesucht ist dann das kleinste k, sodass für einen gegebenen Graphen ein k-lokales Queue/Stack Layout existiert. Im Rahmen von Praxis der Forschung werden der Stand der Forschung, die geplanten Fragestellungen und eigene Vorarbeiten präsentiert.

Communication Optimization by Data Replication for Distributed Graph Algorithms

  • Date: Fri, 27 Sep 2019, 10:00
  • Location: SR 236
  • Speaker: Tobias Ribizel, ITI Sanders; Master's Thesis
  • Inhalt: TBA

Drawing Hypergraphs as Metro Maps

  • Date: Fri, 6 Sep 2019, 14:30
  • Location: SR 301
  • Speaker: Fabian Frank, ITI Wagner; Bachelor's Thesis
  • Inhalt: Eine Metromap-Darstellung eines Hypergraphen ist eine Visualisierung eines Hypergraphen in der jede Hyperkante des Hypergraphen als eine Metrolinie und jeder Knoten als eine Metrostation dargestellt wird. Eine Knotenkreuzung ist ein Knoten, in welchem sich mindestens zwei Metrolinien kreuzen. Ein path-based support Graph G für einen Hypergraphen H ist ein Graph, sodass für jede Hyperkante h von H ein Pfad existiert der in G liegt und genau die Knoten von h enthält. Für einen gegebenen Hypergraphen H = (V, E) und eine gegebene Ordnung jeder Hyperkante zeigen wir, dass es NP-vollständig zu entscheiden ist, ob eine Metromap-Darstellung des Hypergraphen mit höchstens k Kreuzungsknoten existiert, auch wenn die Einbettung bereits vorgegeben ist. Des Weiteren stellen wir einen polynomiellen Algorithmus für den Fall vor, dass die Einbettung vorgegeben und der zugehörige path-based support Graph ein Baum ist.

Engineering Overlapping Community Detection based on the Ego-Splitting Framework

  • Date: Fri, 6 Sep 2019, 14:00
  • Location: SR 301
  • Speaker: Armin Wiebigke, ITI Wagner; Master's Thesis
  • Inhalt: A community is a set of strongly connected nodes in a network. In real-world networks, communities often overlap, meaning that nodes may be part of multiple communities. We evaluate the problem of detecting overlapping communities in a network, using the ego-splitting framework. The ego-splitting framework first searches for the communities of each node in its local neighborhood, the ego-net. Each node is then split into multiple personas to disentangle overlapping communities, each persona corresponding to one of the locally detected communities. The result is a persona graph with non-overlapping communities on which detecting communities is comparatively easy. The ego-splitting framework requires two non-overlapping community detection algorithms, one to analyze the ego-net and one to detect the global communities. While this means that the ego-splitting framework is highly flexible, there is a lack of practical evaluation. We present extensive experimental results for various non-overlapping community detection algorithms and analyze their usefulness in the framework. Our evaluation shows that choosing the right algorithms is essential to detect high quality communities. We extend the ego-splitting framework by three additional steps: We extend the ego-nets, connect the personas of each node, and clean up the detected communities. Our experiments show that our improved version of the ego-splitting framework is able to provide a higher quality than state-of-the-art algorithms on many synthetic graphs. Moreover, the running time of our algorithm is still comparatively low and does not depend on the structure of the network.

Distributed String Sorting Algorithms

  • Date: Wed, 24 Jul 2019, 11:30
  • Location: SR 236
  • Speaker: Matthias Schimek, ITI Sanders; Master's Thesis
  • Inhalt: Although there has been extensive work on sequential and shared-memory parallel string sorting, the problem has not yet been thoroughly studied for distributed systems. In this thesis we present two new distributed string sorting algorithms. Our first algorithm dMSS extends the -- to our knowledge -- only distributed string sorting algorithm with longest common prefix-related optimizations. Furthermore, we present a new approach to compute an ordered partition of a distributed string array such that the number of characters in each set of the partition has about the same value. Our second algorithm dPDSS addresses a major problem of distributed string sorting: the communication of characters which are not required to establish a lexicographical order of the input. In dPDSS distributed bloomfilters are applied to approximately compute the distinguishing prefixes of the input without exchanging the actual strings. Afterwards, the algorithm operates on the (approximate) distinguishing prefixes only. By this means, a significant amount of communication can be saved provided that the distinguishing prefixes are short compared to the strings themselves. Furthermore, we introduce a new string generator producing string data sets with the ratio of the distinguishing prefix length to the entire string length being an input parameter. Our evaluation on up to 1280 processors shows that the presented algorithms clearly outperform their competitors on both generated and real-world data.

Customizable Contraction Hierarchies with Turn Costs

  • Date: Fri, 19 Jul 2019, 14:00
  • Location: SR 301
  • Speaker: Michael Zündorf, ITI Wagner; Bachelor's Thesis
  • Inhalt: Many algorithms are known to solve the shortest path problem on road networks without turn costs. This work focuses on Customizable Contraction Hierarchies and extends this technique to handle networks with turn costs efficiently. We compare different ways to model turn costs in road networks and evaluate which impact those models have on the performance of Customizable Contraction Hierarchies. We also propose some improvements to handle turn costs faster. Albeit most of these improvements were designed for turn costs, some do also apply to Customizable Contraction Hierarchies in general. In addition to that, we introduce a new algorithm to build a Customizable Contraction Hierarchy which is not only simpler but also more efficient than the original algorithm.

A Comparative Study of Overlapping Community Detection Algorithms

  • Date: Wed, 26 Jun 2019, 14:00
  • Location: SR 301
  • Speaker: John Gelhausen, ITI Wagner; Bachelor's Thesis
  • Inhalt: Das Ziel der Erkennung von Communities in Netzwerken ist es, Gruppen von Knoten zu finden welche dicht zueinander und spärlich zu dem Rest des Netzwerkes verbunden sind. Überlappende Communities erlauben es den Knoten Teil mehrerer Communities zu sein. Wir evaluieren neun Algorithmen für die Erkennung von überlappenden Communities, indem wir sie mit Hilfe von Experimenten auf synthetischen Benchmark Netzwerken und realen Netzwerken miteinander vergleichen. Die Algorithmen werden empirisch evaluiert anhand von Performanz-Metriken welche die Ähnlichkeit der gefundenen Communities zu denen der Referenz Communities berechnen. Eine solche Metrik ist zum Beispiel die Normalized Mutual Information (NMI). Wir haben zusätzliche Experimente durchgeführt, um die Algorithmen detaillierter analysieren zu können. Zum Beispiel haben wir überprüft, ob die Algorithmen zu wenige oder zu viele, zu kleine oder zu große Communities finden. Unsere Resultate zeigen, dass OSLOM und MOSES am besten sind, um überlappende Communities zu finden, wobei OSLOM bessere Resultate auf kleinen Netzwerken und MOSES bessere Resultate auf größeren Netzwerken liefert. Unsere Resultate zeigen auch, dass es sehr wichtig ist ergänzende Metriken zu benutzen um die Qualität von den Algorithmen zu evaluieren. Qualitäts-Metriken, wie die NMI oder der Omega Index, berechnen nur die Gesamtqualität eines gefundenen Covers. Dagegen liefern uns ergänzende Metriken mehr Informationen über das Verhalten der Algorithmen. Schlussendlich, während einige Algorithmen gute Resultate auf synthetischen Netzwerken liefern, sind keine der Algorithmen in der Lage die Community Struktur in realen Netzwerken zu erkennen. Das kommt daher, dass die erkannten Communities grundlegend unterschiedlich von den Communities sind die durch Meta-Daten definiert sind.

Effiziente Planung „schöner“ Rundfahrten

  • Date: Tue, 4 Jun 2019, 13:00
  • Location: SR 301
  • Speaker: Sebastian Knapp, ITI Sanders; Master's Thesis
  • Inhalt: There is already a lot of work on calculating the shortest path from a starting point to a destination, and there are also algorithms for finding beautiful round trips. The aim of this work is to combine these two topics and to present algorithms which calculate beautiful round trips considering an intermediate destination and a distance specification. With solving this problem, on the one hand, significantly more beautiful round trips compared to a commercial third-party provider are found and on the other hand, running times of about 100 milliseconds for round trips of a distance of 200 kilometers are achieved using the acceleration techniques of the Contraction Hierachies method. In addition, it is shown that the algorithms also produce qualitatively similar results for longer distances at practical running times.

Transmission Network Expansion Planning for Curing Critical Edges

  • Date: Fri, 17 May 2019, 14:00
  • Location: SR 301
  • Speaker: Lena Winter, ITI Wagner; Masters Thesis
  • Inhalt: In this thesis we discuss the problem to expand transmission networks with the objective to cure critical edges. Critical edges are hereby identified as edges in the transmission network, thats failure would lead to black outs. To identify critical edges we use in this thesis a method proposed by Witthaut et al. [WRZ+16a, WRZ+16b]. This method relies on an graph theoretical approach and can be applied independently from the underlying network model. First we define the problem to expand the transmission network with the objective to cure critical edges accurately. Then we discuss the complexity of the defined problem even for restricted instances of the problem. A major part of this thesis is the development and introduction of different approaches to solve this problem. We will introduce approaches using the underlying network model as well as heuristics solely relying on the graph theoretical simplification of the transmission network. After evaluating the proposed optimization methods we find that the simplification to approach the problem solely graph theoretical leads to large speed ups while retaining the result quality in most cases we tested.

Erweiterungsplanung in elektrischen Netzen mittels dynamischer Programmierung

  • Date: Fri, 12 Apr 2019, 14:30
  • Location: SR 301
  • Speaker: Robert Mumper, ITI Wagner; Bachelor's Thesis
  • Inhalt: Durch die anhaltende Energiewende und den stätig steigenden Strombedarf der Verbraucher sind viele Stromnetze nicht mehr zeitgemäß oder werden in absehbarer Zukunft den Anforderungen nicht mehr entsprechen. Transmission Network Expansion Planning (TNEP) befasst sich unter anderem mit der Ermittlung optimaler Erweiterungen der aktuellen Stromnetze. In dieser Arbeit wird ein dynamischen Programm zur Lösung dieses Optimierungsproblems entwickelt. Da diese Arbeit eine theoretische Betrachtung zur Motivierung des Problems ist, wird lediglich die Klasse der serien-parallelen Graphen betrachtet. Als Referenz wird ein bekanntes Optimierungsprogramm herangezogen, mit dem die Ergebnisse und die Performance verglichen werden. Für alle betrachteten Eingaben liefert der vorgestellte Algorithmus optimale Ergebnisse. Die Laufzeiten des vorgestellten Algorithmus sind außer bei sehr kleinen Stromnetzen schlechter als die des Referenzprogramms.

Algorithmische Werkzeuge zur Analyse von Routingdaten basierend auf historischen Fahrzeugtrajektorien

  • Date: Fri, 12 Apr 2019, 14:00
  • Location: SR 301
  • Speaker: Florian Krone, ITI Wagner; Bachelor's Thesis
  • Inhalt: Korrekte Routing-Services werden immer wichtiger, da sich immer mehr Menschen auf sie verlassen. Daher stellen wir in dieser Arbeit Algorithmen vor, die Routing-Services auf Fehler untersuchen. Die Fehler, die wir finden wollen, sind fehlende Straßen, verbotene Manöver und Abweichungen in der vorhergesagten Fahrtdauer. Dazu nutzen wir historische Fahrzeugtrajektorien. Unser Ansatz ist es, die Trajektorien nachzubilden, indem wir mit den Routingdaten, von Start bis Ziel der Trajektorie, eine Route berechnen. Zur Überprüfung der Nachbildung nutzen wir die Fréchet-Distanz. Durch die Identifizierung der Stellen, an denen die Nachbildung fehlschlägt, können wir mögliche Fehler in den Routingdaten finden. Wir schlagen eine Clustering-Strategie vor, um Punkte herauszufiltern, an denen die Nachbildung durch GPS-Ausreißer oder irrationales Verhalten des Fahrers nicht möglich ist. Dadurch identifizieren wir Punkte, an denen Fehler in den Routingdaten wahrscheinlich sind. Zusätzlich haben wir auf Basis der Zwischenergebnisse unseres Algorithmus Maßzahlen entwickelt, die einen Vergleich der Qualität verschiedener Routing-Services ermöglichen.

A Brief Introduction to Cycle Bases and Matroids

  • Date: Wed, 10 Apr 2019, 14:00
  • Location: SR 301
  • Speaker: Torsten Ueckerdt, ITI Wagner
  • Inhalt: Roughly speaking, a cycle basis of a graph G is a set of cycles that generates all cycles in G; an important case being a generation in terms of linear combinations over a field Q. Roughly speaking, a matroid of a ground set U is a set of subsets of U that is downwards-closed and always allows greedy extension to a maximum size subset; an important case being linear independent vectors in a vector field, as well as, acyclic subgraphs of a graph. In this talk, we seek to give a gentle introduction to both, cycle bases and matroids, trying to highlight some of their interconnections.

Concurrent Dynamic Quotient Filters - Packing fingerprints into atomics

  • Date: Tue, 9 Apr 2019, 14:30
  • Location: SR 301
  • Speaker: Robert Williger, ITI Sanders; Master's Thesis
  • Inhalt: Quotient filters are approximate membership data structures that can be used for a variety of applications, from speeding up database accesses to uses in computational biology. We look at several improvements to common quotient filters which are orthogonal to each other and can be combined as needed. We show how to use compact arbitrary length data types to reduce the memory usage by up to seven times and get very close to the theoretical optimum. We also present two different approaches to building concurrent quotient filters that use localized locking or no locking at all. This leads to a good linear speedup with increasing thread counts and an increase in performance of up to four times compared to traditional locking techniques. Additionally, we describe how to use a multilevel quotient filter structure to implement dynamic growing while still allowing for a user defined maximum false positive rate and being less than 10% slower compared to nongrowing variants.

Kompressionstechniken für Beschreibungen von SAT Formeln

  • Date: Tue, 9 Apr 2019, 14:00
  • Location: SR 301
  • Speaker: Jens Manig, ITI Sanders; Bachelor's Thesis
  • Inhalt: In dieser Arbeit werden verschiedene Ansätze zur Kompression des DIMACS-Dateiformats für aussagenlogische Formeln vorgestellt. Das DIMACS-Dateiformat findet beim SAT-Solving Verwendung und enthält logische Klauseln welche als Zahlen repräsentiert, aber als Strings gespeichert werden. Daher beschäftigt sich die vorliegende Arbeit damit, Vorwissen über dieses Format für die Kompression zu verwenden. Der erste entwickelte Ansatz ist eine Umwandlung der Strings in Zahlen mit einer festen Anzahl Bytes pro Zahl, der zweite Ansatz extrahiert die Vorzeichen in Bitvektoren und nutzt eine variable Byteanzahl pro Zahl, der dritte nutzt einen Move-to-Front Ansatz, der vierte erreicht zusätzliche Kompression über Präfixmatching aufeinanderfolgender Zeilen und der letzte Ansatz ist eine Kombination aus Move-to-Front und Präfixmatching. Von den vorgestellten Ansätzen ist die Kombination die beste bezüglich des Kompressionsfaktors. Bei der Laufzeit ähneln sich die entwickelte Ansätze stark, sowohl in der Kodierung als auch Dekodierung. Im Vergleich zum allgemeinen ZIP-Kompressionsverfahren ist die Kodierungszeit bedingt besser, während die Dekodierungszeiten von ZIP noch deutlich geringer sind. Beim Kompressionsfaktor ist die Kombination nicht weit von ZIP entfernt, und eine Kombination aus ZIP und den vorgestellten Verfahren führt zu den insgesamt besten Kompressionsraten.

Algorithms for the Pagination Problem on Public Transit Networks

  • Date: Fri, 5 Apr 2019, 14:00
  • Location: SR 301
  • Speaker: Moritz Halm, ITI Wagner; Bachelor's Thesis
  • Inhalt: Diese Arbeit beschäftigt sich mit der Frage, wie Reiseprofile auf öffentlichen Verkehrsnetzwerken unter Einsatz von Paginierung berechnet werden können. Paginierung bedeutet hier das Aufteilen eines Profiles in Abschnitte (Seiten), die nacheinander berechnet und ausgegeben werden. Dabei hat die Sortierung der Reisen innerhalb eines Profils entscheidenden Einfluss auf die Nützlichkeit des Ergebnisses. Wir vergleichen Sortierungen von Reisen nach Ankunfts- und Abfahrtszeit, und schlagen außerdem vor, Reisen nach dem frühesten Zeitpunkt, von dem an sie optimal sind, zu sortieren. Wir stellen verschiedene auf dem RAPTOR-Algorithmus basierende Ansätze vor, wie Reisen in einer jeder dieser Sortierungen entsprechenden Reihenfolge berechnet werden können. Zunächst beschreiben wir, wie der von Wagner und Zündorf (2017) vorgestellte Profilalgorithmus jede dieser Sortierungen liefert, wenn man die Reihenfolge der durchgeführten Vorwärts- und Rückwärtssuchen in geeigneter Weise anpasst. Speziell für die Sortierung nach Ankunftszeit, können allerdings kürzere Laufzeiten durch den Einsatz von rRAPTOR erreicht werden, indem man den Algorithmus auf einem umgekehrten Netzwerk ausführt. Für die anderen beiden Sortierungen erreichten wir leichte Geschwindigkeitsvorteile durch eine Kombination beider Ansätze. Wir haben die Laufzeit unserer Algorithmen anhand von Experimenten auf dem öffentlichen Verkehrsnetzwerk der Schweiz gemessen.

On Rigidity of Graphs

  • Date: Mon, 1 Apr 2019, 14:00
  • Location: SR 301
  • Speaker: Oliver Suchan, ITI Wagner; Bachelors Thesis
  • Inhalt: A rigid framework refers to an embedding of a graph, in which each edge is represented by a straight line. Additionally, the only continuous displacements of the vertices, that maintain the lengths of all edges, are isometries. If the edge lengths are allowed to only be perturbed in first order, the framework is called infinitesimally rigid. In this thesis the degrees of freedom of a certain, new type of rigidity, called edge-x-rigidity, for a given framework in the 2-dimensional Euclidean space are examined. It does not operate on the length of an edge, but on the intersection of an edge's support line and the x-axis. This may then help to determine the class of graphs where every edge intersection with the x-axis can be repositioned along the x-axis, such that there still exists a framework that satisfies these position constraints

Fall 2018

A specialized model for optimized memory consumption of time-dependent Contraction Hierarchies

  • Date: Tue, 26 Mar 2019, 14:00
  • Location: SR 301
  • Speaker: Sebastian Schmidt
  • Inhalt: Zeitabhängige Routenplanung wird momentan praktikabel, da mehr und mehr GPS-Daten über Autofahrten verfügbar werden. Aber die Nutzung von detaillierteren Daten für Zeitabhängige Routenplanung führt auch zu höherem Speicherverbrauch für Beschleunigungstechniken. Wir optimieren den Speicherverbrauch von Zeitabhängigen Customizable Contraction Hierarchies. Dazu stellen wir ein Optimierungsschema für die Eingabefunktionen vor, um den Speicherverbrauch der exakten TDCCH auf den optimierten Eingabedaten zu verringern. Wir argumentieren, dass unser Gewichtsoptimierungsschema nur sehr kleine signifikante Fehler einfügt, und größtenteils unter dem Rauschen in den Eingabedaten bleibt.

Campus-Routenplaner

  • Date: Fri, 22 Mar 2019, 14:00
  • Location: SR 301
  • Speaker: -, ITI Wagner / PSE-Abschlussvorträge
  • Inhalt: Das Thema der PSE-Gruppen ist der Campus-Routenplaner.

Recognizing Planar Laman Graphs

  • Date: Fri, 22 Feb 2019, 14:00
  • Location: SR 301
  • Speaker: Jonathan Rollin, ITI Wagner; Guest from the FernUni Hagen
  • Inhalt: Laman graphs are the minimally rigid graphs in the plane and can be characterized in geometric as well as combinatorial ways. We present two algorithms for recognizing planar Laman graphs. A simple algorithm with running time O( n^(3/2) ) and another one with running time O(n log^3 n) based on latest planar network flow algorithms. Both improve upon the previously fastest algorithm for general graphs by Gabow and Westermann [Algorithmica, 7(5-6):465--497, 1992] with running time O(n (n log n)^(1/2) ). Joint work with Lena Schlipf and André Schulz.

Ridesharing with Multiple Riders

  • Date: Fri, 15 Feb 2019, 14:00
  • Location: SR 301
  • Speaker: Oliver Plate, ITI Wagner; Masters Thesis
  • Inhalt: Ridesharing bezeichnet die gemeinsame Nutzung eines Fahrzeuges für den Transport von Personen. Die wesentlichen Ziele dabei sind ökonomisch (Kosten teilen) als auch ökologisch (Reduzieren der Umweltbelastung). Das Thema meiner Arbeit ist die Realisierung von Mitfahrgelegenheiten, indem Anfragen zu den vorhandenen Angeboten zugeordnet werden. Der Themenkreis ist bereits in vielen Arbeiten mit unterschiedlichen Schwerpunkten untersucht worden, dazu geben wir einen Überblick mit diversen Klassifikationen, Optimierungszielen und Lösungsansätzen. Die vorliegende Arbeit befasst zuerst mit der Version, in der ein Fahrer Mitfahrgelegenheiten für eine oder mehrere Personen entlang einer von ihm zu fahrenden Route anbietet, wobei gewisse Umwege in Kauf genommen werden. Dazu wird ein Algorithmus vorgestellt, der auf einer schnellen und dynamischen Lösung für one-to-many (many-to-one) Kürzeste-Wege-Anfragen aufbaut und somit rechenintensive einzelne Kürzeste-Wege-Anfragen umgeht. Dabei wird die Bestimmung kürzester Pfade mithilfe einer Kombination aus Bucket-Einträgen und Contraction Hierarchies durchgeführt. Zudem werden, mit Blick auf die Antwortzeiten von Anfragen, zusätzliche Strategien zur Beschleunigung eingeführt. Die Flexibilität unseres Lösungsansatzes wird gezeigt, indem mit geringer Abänderungen auch das sogenannte Taxi-Ridesharing-Problem angegangen wird. Verschiedene Messungen hinsichtlich Antwortzeiten, Trefferquoten und der Qualität der Treffer dienen dazu, die Qualität der erstellten Lösung zu untersuchen und Vergleiche mit anderen Lösungsverfahren anzustellen.

Collinearity of Independent Sets - Algorithmic Challenges

  • Date: Mon, 14 Jan 2019, 14:00
  • Location: SR 301
  • Speaker: Almut Demel, ITI Wagner; Masters Thesis
  • Inhalt: A subset $S \subset V(G)$ of vertices of a planar graph G is called collinear if G admits a plane straight-line drawing where all the vertices in S lie on a line. The aim of this thesis is to find a lower bound for the maximal number of collinear vertices in graphs of a certain graph class. Our approach is to examine whether independence or a fixed minimal distance between vertices in a subset of vertices is a sufficient condition for collinearity. An independent set of vertices with minimal pairwise distance d is called d-scattered. The size of maximum independent sets is linear in the number of vertices in planar graphs. However, in general planar graphs the number of collinear vertices is limited by a sublinear upper bound. Thus we examine specific graph classes, in particular series-parallel graphs and 4-connected triangulations. For the former we show that any fixed minimal pairwise distance between vertices of an independent set is not a sufficient condition for collinearity. Though we introduce a meaningful subclass of series-parallel graphs, in which 3-scattered sets are always collinear. For 4-connected triangulations we show that independent sets are in general not collinear and present forbidden substructures. We further present approaches to prove collinearity of d-scattered sets and outline the limits of these approaches.

Route Planning with Temporary Road Closures

  • Date: Fri, 21 Dec 2018, 14:00
  • Location: SR 301
  • Speaker: Christian Bräuer, ITI Wagner; Masters Thesis
  • Inhalt: We study the problem of route planning with temporary road-closures. In practice, road segments, e.g. in residential or downtown areas, are closed at certain times. In addition, in many countries general weekend and night driving bans apply. Mainly trucks are affected by these driving bans. The road-closures may inflict waiting time along a route for which parking lots are needed. Routes with mathematical earliest arrival at the destination may contain loops and other detours. In this thesis we define a model which yields reasonable routes in practice, where the road-closures are taken into account. We distinguish between the travel time and the driving time of a route. The travel time is the time span between the start and the end of the route. The driving time is the duration in which the driver moves his vehicle. The distinction between travel time and driving time may yield multiple routes per route planning query. Thus, we use the concept of Pareto-optimality. We present an algorithm which calculates Pareto-optimal routes according to the model and adapt known speed up techniques to the algorithm. Using heuristic algorithms, we achieve a running time per query of a few seconds.

Dynamische Baumkompression mit Top Trees

  • Date: Wed, 19 Dec 2018, 13:30
  • Location: To appear.
  • Speaker: Jan Ellmers, ITI Sanders; Bachelors Thesis
  • Inhalt: In dieser Arbeit sehen wir uns Verbesserungsmöglichkeiten für Tree Compression mit Top Trees an. Wir entwickeln zwei neue Navigationsfunktionen, die ebenfalls in logarithmischer Zeit laufen und keine zusätzlichen Informationen benötigen, und zeigen, wie die Navigation von Hübschle-Schneider and Raman mit logarithmischem Platz realisiert werden kann. Des Weiteren zeigen wir, wie wir weniger Information für den Top Tree speichern müssen, und stellen neue Regeln für das Erstellen des Top Trees vor, die bessere Kompressionsraten und Laufzeiten haben. Zudem untersuchen wir die Möglichkeit, an einem komprimierten Baum Änderungen vorzunehmen. Wir testen unsere Ergebnisse an dem XML Corpus von Hübschle-Schneider and Raman und schauen uns an diesem die praktischen Auswirkungen von Slowing Down an.

Optimierung globaler Flexibilität von elektrischen Lasten mittels Genetischer Algorithmen

  • Date: Fri, 14 Dec 2018, 14:30
  • Location: SR 301
  • Speaker: Johanna Rey, ITI Wagner; Bachelors Thesis
  • Inhalt: Lösungsansätze um mit Demand Side Management auf Veränderungen durch die Energiewende zu reagieren sind größtenteils auf private Haushalte fokusiert. Dies führt dazu, dass in der Industrie Demand Side Management kaum eingesetzt wird. Durch größere Mengen an elektrischen Lasten die in der Industrie verwendet werden, liegt jedoch gerade dort ein großes Potential in der Anwendung liegt. Wir entwickeln einen Genetischen Algorithmus, der mit dem von Barth et al. entwickelten neuen Scheduling-Ansatz berechnet was mit unterschiedlichen Graden an Flexibilität erreicht werden kann. Relativ zu den durch ein MIP-Solving errechneten Werten erreichen wir mit unserem Algortihmus Ergebnisse, die innerhalb von vier Sekunden Werte mit im Schnitt 95% der Qualität aufweisen können.

A Flow-Based Initial Partitioning Algorithm for Multilevel Graph Partitioning

  • Date: Fri, 14 Dec 2018, 14:00
  • Location: SR 301
  • Speaker: Tim Niklas Uhl, ITI Wagner; Bachelors Thesis
  • Inhalt: The graph partitioning problem is to partition the nodes of a graph into k blocks of roughly equal size such that the number of edges running between blocks is minimized. Most state-of-the-art graph partitioning algorithms are based on the multilevel scheme which obtains a partition of a graph by recursively contracting the input graph, initially partitioning the contracted graph and refining the partition while undoing the contraction. While flow-based approaches have been used for partition refinement, flow-based initial partitioning has not received any attention. In this thesis we investigate a flow-based initial partitioning algorithm. We incorporated node and edge weight into the bipartitioning algorithm FlowCutter. FlowCutter computes a sequence of s-t-cuts of increasing balance using incremental flow augmentations. We evaluate the performance of the adapted algorithm by replacing the initial partitioner in the multilevel graph partitioner KaHiP. We show that our algorithm yields better initial partitions than KaHiP when contraction is stopped earlier and is also able to improve the final cut size. On 16 out of 30 graphs from the Walshaw Benchmark set we find cuts with the same size as KaHiP, while we achieve an improvement on 5 graphs. We further show that coarsening based on the Louvain method is able to improve cut capacities of the initial partition. Using our algorithm for initial partitioning we find better initial partitions than KaHiP for 24 out of 30 graphs even when KaHiP stops contraction early and the resulting graphs are therefore larger than those obtained using Louvain.

Distributed Kernelization for Independent Sets

  • Date: Tue, 11 Dec 2018, 13:00
  • Location: SR 301
  • Speaker: Tom George, ITI Sanders; Bachelors Thesis
  • Inhalt: To appear.

A Practical Analysis of Kernelization Techniques for the Maximum Cut Problem

  • Date: Fri, 30 Nov 2018, 13:00
  • Location: SR 236
  • Speaker: Damir Ferizovic, ITI Sanders; Masters Thesis
  • Inhalt: In this work we study existing and new kernelization techniques for a famous NP-Complete problem: extsc{Max-Cut}. Primarily the unweighted case is being studied, but also the signed and weighted versions are addressed to some degree. While the unweighted extsc{Max-Cut} problem is not ubiquitous in practice, the signed and weighted version are. Furthermore, the achieved results from the unweighted case do set a lower bound for the expectations to be had on the more general cases. At the beginning we provide a broad overview of existing techniques and the fixed-parameter tractable algorithms they imply. We then introduce any new set of rules that have been created within our work and analyze the used reduction rules from previous works. Moreover, we outline our approach on how we integrated the reductions into our testing suite. For some set of rules a proof is also given if they are a subset of another one; making that other rule be able to do everything it can do, and more. Thus, this allows us to ignore such rules from further regard. Within the closure of our work, a practical analysis is given: The usefulness of each reduction rule is scrutinized and the results of testing a wide range of graphs is provided. We also specifically analyze the practical results implied by the work from Etscheid and Mnich cite{crowston2014satisfying}, which guarantee a linear sized kernel. As utilized in their work, we also deploy related techniques  to compute a polynomial-time solvable subgraph for further reduction of the kernel size.

Approximate Transit Node Routing - An Approximate Distance Oracle for Social Networks

  • Date: Tue, 27 Nov 2018, 14:00
  • Location: SR 211
  • Speaker: Katharina Flügel, ITI Sanders; Bachelors Thesis
  • Inhalt: Computing shortest paths is a fundamental graph problem with many real-life applications. Naive methods such as Dijkstra's algorithm are not fast enough for today's graphs with millions of vertices and edges. We present approximate transit node routing (aTNR), an approximate distance oracle for social networks with a limited additive error. It is an extension to the transit node routing framework which was originally developed for road networks. Distances can be computed significantly faster with aTNR compared to the naive approaches. We can adjust the maximum error for each graph and even compute exact distances at the cost of a longer preprocessing time. The distance between two vertices is computed using a three-hop via the transit nodes, a small set of central vertices. For each vertex, a set of access nodes is chosen from the transit nodes via which the three-hop is computed for that vertex. We evaluate aTNR on multiple different graphs regarding the preprocessing time, preprocessing size, query time, and approximation error. Smaller graphs are preprocessed in mere seconds, and even graphs with more than one hundred million edges can be preprocessed in less than two minutes. Most queries take about 20 microseconds and cause only a very small approximation error. Our distance oracle is particularly proficient in terms of space consumption as it requires less than 50 bytes per vertex. In addition to aTNR, we define the Voronoi search distance oracle, a variation of aTNR which increases the approximate error in exchange for potentially faster preprocessing and queries.
  • Date: Fri, 23 Nov 2018, 14:00
  • Location: SR 301
  • Speaker: Kai Fieger, ITI Sanders; Masters Thesis
  • Inhalt: This thesis presents a parallel algorithm that solves the longest path problem for weighted undirected graphs. The algorithm uses dynamic programming and graph partitioning in order to find the longest path between two vertices. The algorithm finds the optimal longest path and not an approximation of it like many other approaches for NP-complete problems. We first present the serial algorithm and then how it can be parallelized. Through experiments we measured the speedup of the parallel algorithm compared to its serial counterpart. We achieved reasonable speedups. We also demonstrate that the algorithm has possible applications in solving the Hamiltonian cycle problem.

A Block-Cut-Tree-based Switching Algorithm for Cacti

  • Date: Tue, 20 Nov 2018, 14:00
  • Location: SR 301
  • Speaker: Florian Krüger, ITI Wagner; Bachelors Thesis
  • Inhalt: The Maximum Transmission Switching Flow Problem (MTSF) focuses on maximizing the power flow within an electrical power grid by allowing transmission lines to be removed from the grid, where the latter procedure is referred to as switching. MTSF is an optimization problem and known to be NP-complete, even on strongly restricted graph classes such as cacti. Therefore the standard approach is to solve an instance of its formulation as a Mixed Integer Linear Program (MILP). In this thesis, we present an algorithm of exponential complexity to solve MTSF for unbounded cacti grids to optimality by utilizing the structure of its Block-Cut-tree. We achieve this by solving a power flow model for subgraphs and then composing partial solutions to an overall solution. We furthermore introduce a heuristic to gain improvements in terms of running time. In conclusion we evaluate the results of our algorithm in comparison to solving the MILP on real power grid instances, that have been modified to represent cacti. We identify the cause for the exponential running time and also assess the solution deviation from optimality of the heuristic.

PoI Anfragen im Vergleich

  • Date: Tue, 20 Nov 2018, 13:30
  • Location: SR 301
  • Speaker: Hamidulah Doust, ITI Wagner; Bachelors Thesis
  • Inhalt: Fast täglich benutzt man einen Navigator um den schnellsten Weg von einem Startpunkt zu einem Zielpunkt zu berechnen. Dabei existieren Anfragen, die generell öfters verwendet werden. Beispielsweise Interessiert man sich für die Haltestellen des öffentlichen Verkehrs oder die nächste Ladestation für Elektrofahrzeuge. Um diese Art von Anfrage zu lösen benutzt man kürzeste Wege Algorithmen. Speziell interessant in dieser Arbeit ist das "One to Many" Problem, sprich von einem Startpunkt die kürzesten Wege zu einer Menge von Zielpunkten zu bestimmen. Im Laufe dieser Bachelorarbeit vergleichen wir unterschiedliche "One to Many" Algorithmen. Ebenfalls werden neue "One to Many" Algorithmen vorgestellt. Wir testen und evaluieren diese Algorithmen auf verschiedene Datensätzen, u.a. auf Haltestellen des öffentlichen Verkehrs als auch Ladestationen für Elektrofahrzeuge.

Dijkstra-basiertes Mapmatching

  • Date: Tue, 20 Nov 2018, 13:00
  • Location: SR 301
  • Speaker: Peter Maucher, ITI Wagner; Bachelors Thesis
  • Inhalt: In dieser Arbeit beschreiben wir einen neuen Ansatz zur Lösung des Mapmatching-Problems basierend auf Dijkstras Algorithmus. Zu einem sich in einem Straßennetz- werk bewegenden Objekt werden Positionsmessdaten aufgezeichnet. Im Mapmatching-Problem wird zu diesen Messdaten ein möglichst plausibler Pfad in dem Graphen gesucht, der das Straßennetzwerk repräsentiert. Dijkstras Algorithmus löst das kürzeste-Wege-Problem effizient. Unser Ansatz erweitert diesen Algorithmus, indem die Kantengewichte abhängig von den gemessenen Daten verringert werden. Mit den geänderten Kantengewichten entsteht eine Lösung des Mapmatching-Problems. Zur Anpassung der Gewichte werden Wegintegrale über die Graphkante verwendet. Wir zeigen die Effizienz unseres Algorithmus auf echten Daten, indem wir ihn mit einem unmodifizierten Dijkstra vergleichen. Die Genauigkeit wird durch den Vergleich mit einem Standardalgorithmus für Mapmatching mit Namen ST-Matching gezeigt.
  • Date: Mon, 12 Nov 2018, 14:00
  • Location: SR 236
  • Speaker: Guangping Li, ITI Sanders; Masters Thesis
  • Inhalt: Stochastic local search (SLS) is an elementary technique for solving combinational problems. In the first section of this paper, we introduce an efficient SLS heuristic solver for Boolean Satisfiability Problem (SAT), in which the decisions only based on the probability distribution. We experimentally evaluate and analyze the performance of our solver in a combination of different techniques, including simulated annealing and walkSAT. With formula partition, we introduce a parallel version of our solver in the second section. The parallelism improves the efficiency of the solver. Using different random generator in solving the sub-formula can bring further improvement in performance to our parallel solver.

Kernelization for Maximum Weight Independent Sets and Distributed Connected Components

  • Date: Fri, 9 Nov 2018, 14:00
  • Location: SR 301
  • Speaker: Sebastian Lamm, ITI Sanders - Research Overview
  • Inhalt: After spending half a year at the TAA research group in Vienna, I would like to present an overview of the work I did during my stay. The talk will most likely be split into two parts. First, I will talk about our recent results on kernelization for the maximum weight independent set problem. In particular, I will present a set of novel reduction rules that are able to significantly reduce the size of large real-world graphs. This allows us to compose an efficient branch-and-reduce algorithm that outperforms the current state-of-the-art. Second, we will take a look at distributed algorithms for finding connected components. My work here focuses on achieving scalability by focusing on communication efficieny. For this purpose, I developed and engineered an algorithm that uses various types of local contraction routines. In addition to providing state-of-the-art performance, the optimizations deployed can be used to significantly speedup existing approaches.

Parallelization Techniques for Customizable Contraction Hierarchies

  • Date: Fri, 2 Nov 2018, 13:30
  • Location: SR 301
  • Speaker: Max Göttlicher, ITI Wagner; Bachelors Thesis
  • Inhalt: Mit Customizable Contraction Hierarchies können sehr schnelle Routenabfragen realisiert werden, wie sie für die Echtzeitvorschau in einigen Onlinediensten benötigt werden. Darüber hinaus können auch aktuelle Verkehrsdaten verwendet werden, um eine realistische Fahrzeit zu erhalten. Ermöglicht wird dies durch eine zweiphasige Vorberechnung, die zunächst eine metrikunabhängige Graphstruktur berechnet, die erst in der zweiten Phase um eine Metrik erweitert wird. Die zweite Phase bietet eine guten Ansatzpuntk für weitergehende Optimierungen, da sie im Vergleich zur ersten Phase relativ oft ausgeführt werden muss und eine lange Vorberechnungszeit die Verwendung aktueller Verkehrsdaten verhindern würde. Durch Parallelisierung kann die Laufzeit stark reduziert werden. Zusätzliche Beschleunigung kann durch verschiedene sequentielle Optimierungen erreicht werden, die auch bei einer parallelen Ausführung noch zum Tragen kommen.

Effizienter Umlegungsalgorithmus für den öffentlichen Verkehr mit Fahrzeugkapazitäten

  • Date: Fri, 2 Nov 2018, 13:00
  • Location: SR 301
  • Speaker: Robin Berger, ITI Wagner; Bachelors Thesis
  • Inhalt: Ein wichtiges Problem bei der Verkehrsplanung im öffentlichen Verkehr ist es, für einen gegebenen Fahrplan und eine gegebene Nachfrage zu wissen, wie viele Personen welches Fahrzeug verwenden. In dieser Arbeit wird ein Verfahren vorgestellt, das dieses Problem löst. Hierbei werden auch Fahrzeugkapazitäten und die Tatsache betrachtet, dass Fahrgäste versuchen überfüllte Fahrzeuge zu meiden.

Relaxed Priority Queues: Using buffers to guarantee quality

  • Date: Wed, 31 Oct 2018, 11:30
  • Location: SR 236
  • Speaker: Holger Ebhart, ITI Sanders; Masters Thesis
  • Inhalt: In this work we present a new approach for a concurrent relaxed priority queue design. Due to popularity of multi-core systems we see a need to adopt basic data structure to these systems for further performance increase. We present the buffered multi queue, a priority queue adopted for concurrent access. The buffered multi queue supports fast insertions into per-thread local priority queues and still guarantees a low rank error for removals. To hold guarantees we propose an output buffer where small elements were moved into. Elements were only popped from this buffer. By adjusting the capacity of the output buffer, we have a trade-off between high throughput rates and good quality. Relaxation of the queue is introduced via the buffer and allows us to return only one of the smallest elements. For this behavior we use a more relaxed semantic than for sequential priority queues, as such a semantic becomes necessary when accessing a priority queue concurrently. We prove that the maximal rank error is in O(buffer_size) as well as the maximal starvation. The starvation is a newly introduced quality metric, stating the time a queued element is skipped for output. Within experiments with different workloads and against competitors we show that our approach keeps up with others in performance. But in quality of returned elements we outperform all other approaches.

Local Page Numbers

  • Date: Fri, 12 Oct 2018, 14:30
  • Location: SR 148
  • Speaker: Laura Merker, ITI Wagner; Bachelors Thesis
  • Inhalt: Ein Buch besteht aus Halbebenen, sogenannten Seiten, die eine gemeinsame Gerade haben, auch Spine genannt. Ein Graph wird in ein Buch eingebettet, indem die Knoten auf dem Spine angeordnet werden und jede Kante auf genau einer Seite eingebettet wird, wobei sich Kanten auf der gleichen Seite nicht kreuzen dürfen. Eine Bucheinbettung heißt k-lokal, wenn an jedem Knoten maximal k Seiten beteiligt sind. Die (global) Page Number ist die Mindestanzahl an Seiten, die für eine Bucheinbettung benötigt wird, während die local Page Number das kleinste k ist, für das es eine k-lokale Bucheinbettung gibt. Wir finden Schranken für die local Page Number einiger Graphklassen wie vollständige Graphen, planare Graphen und k-Bäume. Im Anschluss betrachten wir das NP-schwere Problem, eine k-lokale Bucheinbettung für einen gegebenen Graphen mit fester Knotenreihenfolge zu finden.

Top Trumps - The Graph-Theoretical Properties of a Children's Game

  • Date: Fri, 12 Oct 2018, 14:00
  • Location: SR 148
  • Speaker: Lasse Wulf, ITI Wagner; Masters Thesis
  • Inhalt: The game "Top Trumps" is a simple card game played between two players. Each Top Trumps deck has a certain theme to it, say for example the theme "sports cars". Then, on each card, there is a picture of a certain sports car and a set of categories describing the car, e.g. "top speed", "engine power", "engine displacement" and "weight". One player begins and chooses one of the categories, the players compare the values of the top two cards of their decks in the chosen category. The winner shuffles the two cards to the bottom of his or her deck and is choosing player for the next round. Although the game is quite simple and has never been investigated scientifically, the surrounding theory is surprisingly elegant and non-trivial, especially from a graph-theoretical point of view. In this talk, we are concerned with optimal play, fairness of the game, and first-player-advantage. We also see that the topic is strongly connected to the concept of the width of a partially ordered set. In particular, we show a generalization of the well-known Dilworth's theorem in terms of "Top Trumps deck realizations".

Verkabelung von Windfarmen auf Bäumen

  • Date: Fri, 5 Oct 2018, 14:00
  • Location: SR 236
  • Speaker: Dominik Stampa, ITI Wagner; Bachelors Thesis
  • Inhalt: The wind farm cabling problem deals with installing cables between wind turbines and substations of a wind farm in order to conduct electric power from turbines to substations. For this task several types of cables are given and the cheapest cable installation is to be found. In the model, turbines and substations are represented by vertices of a graph. Cables can only be placed at the edges of this graph. Whereas this problem is NP-hard on general graphs, in this thesis the wind farm cabling problem is considered on tree graphs only. We describe different variants of the problem with the help of several parameters and analyse their complexity. Amongst others, a polynomial-time algorithm for the wind farm cabling problem on path graphs, where the produced power of a wind turbine has to be conducted to exactly one substation, is given. On trees the cabling problem is already NP-hard if an arbitrary set of cable types is given - no matter how the values of other parameters are chosen.

Spring 2018

Wind Farm Cabling using Spectral Clustering

  • Date: Tue, 18 Sep 2018, 14:30
  • Location: SR 301
  • Speaker: Niklas Fuhrberg, ITI Wagner; Bachelors Thesis

Drawing a Level Planar Graph with Fixed Slopes

  • Date: Tue, 18 Sep 2018, 14:00
  • Location: SR 301
  • Speaker: Nadine Krisam, ITI Wagner; Bachelors Thesis
  • Inhalt: In this thesis we study level planar drawings with two fixed slopes of the edges (LP2-drawings). We present an algorithm that, given a level planar graph, decides whether it has an LP2-drawing with an additional requirement of rectangular inner faces. After that, we drop the requirement of rectangular inner faces and provide an algorithm that decides whether a general LP2-drawing exists. Both algorithms have polynomial running time. In the conclusion, we give an outlook how to extend the latter algorithm to work on multiple fixed slopes.

Parallel Greedy Graph Matching

  • Date: Thu, 6 Sep 2018, 14:00
  • Location: SR 301
  • Speaker: Fredrik Manne, ITI Sanders; Guest from the University of Bergen, Norway (Department of Informatics)
  • Inhalt: The weighted matching problem has numerous applications in scientific computing. Although there exists efficient polynomial time algorithms for the problem, in practise these are often too slow for large problems. It is therefore of interest to study fast approximation algorithms. In this presentation we will look at an approximation algorithm called the Suitor algorithm that computes the same solution as the greedy approach, but does so without sorting, and that is also higly amendable to parallelization. From this algorithm we branch out in various directions to investigate how it can be applied to different matching problems. Fredrik Manne is a professor at the University of Bergen and is currently on sabatical leave at KIT. His main interests are in combinatorial scientific computing and in particular in designing and implementing parallel algorithms.

Hierarchical Task Network Planning using SAT Techniques

  • Date: Tue, 24 Jul 2018, 13:00
  • Location: SR 301
  • Speaker: Dominik Schreiber, ITI Sanders; Masters Thesis
  • Inhalt: Automated planning is useful for a wide range of general decision-making processes in the area of Artificial Intelligence. The solving approach of encoding a planning problem into propositional logic and finding a solution with a SAT solver is a well established method. Likewise, a planning model called Hierarchical Task Networks (HTN) which enhances planning problems with expert knowledge can help to find structured plans more easily and efficiently. This work focuses on combining these techniques by using SAT solving techniques to resolve HTN planning problems. Initially, a previous approach to encode HTN problems in SAT by (Mali and Kambhampati, 1998) is analyzed, and various shortcomings are identified from today's perspective. Then, three original encodings are proposed: Grammar Constrained Tasks (GCT) which is inspired by one of the previous encodings and is the first to feature modern HTN domains and recursive task relationships; Stack Machine Simulation (SMS) which is designed for incremental SAT solving and works reliably on all special cases; and Tree-like Reduction Exploration (T-REX), which leads to a particularly efficient solving process due to its short amount of needed iterations and various introduced optimizations. All encodings are implemented to exploit existing HTN grounding routines, and the T-REX approach features a novel abstract formula notation and an efficient Interpreter application tailored to the encoding. In addition, an Anytime plan length optimization within T-REX is proposed. Experiments show that SMS dominates GCT, and T-REX dominates SMS. The proposed T-REX planning framework outperforms a state-of-the-art classical SAT planner on various domains. Regarding run times, T-REX cannot compete with state-of-the-art HTN planners yet, but is still an appealing planning approach due to its robustness, its plan length optimization, and its proving abilities. By the design, implementation and evaluation of T-REX, the work at hand demonstrates that HTN planning via SAT Solving is a viable option and worthy of the attention of future research.

Support Vector Machines via Multilevel Label Propagation

  • Date: Wed, 11 Jul 2018, 13:00
  • Location: SR 301
  • Speaker: Matthias Schmitt, ITI Sanders; Bachelors Thesis
  • Inhalt: The time complexity of support vector machines (SVM) prohibits training on huge data sets with millions of samples. Multilevel SVMs were developed to allow for time efficient training on huge data sets. In this thesis we present a new faster multilevel support vector machine framework which utilizes the near-linear time label propagation algorithm for the construction of the problem hierarchy. While regular SVM perform the entire training in one - time consuming - optimization step, multilevel SVMs first build a hierarchy of problems decreasing in size that resemble the original problem and then train an SVM model for each hierarchy level benefiting from the solved models of previous levels. Our experiments on a large number of benchmark data sets show that our framework (KaMLSVM), when compared to the previous best multilevel SVM, achieves a speed-up of more that two on almost all data stet and up to 20 on large data sets with many features.

"Quadratic"-work construction of terabyte-scale text indexes

  • Date: Wed, 4 Jul 2018, 13:00
  • Location: SR 301
  • Speaker: Juha Kärkkäinen, ITI Sanders; Guest from the University of Helsinki (Department of Computer Science)
  • Inhalt: Suffix array and LCP array are fundamental data structures underlying many modern text indexes, and their construction is often a main bottleneck in index construction. In this talk, we review some recent practical external memory algorithms for constructing these data structures even for terabyte-scale texts. The algorithms share a common approach that results in work and I/O complexities involving the term n^2/M, where n and M are the sizes of the input and the main memory, respectively. Such quadratic complexities are easy to dismiss as unscalable but we show experimental results and theoretical arguments supporting the practicality of this approach.

Partial Contraction Hierarchies

  • Date: Fri, 22 Jun 2018, 14:30
  • Location: SR 301
  • Speaker: Jonas Meier, ITI Wagner; Bachelors Thesis
  • Inhalt: In großen oder komplexen Straßengraphen können Kürzeste-Wege-Anfragen mit naivem Dijkstra-Algorithmus sehr schnell sehr langsam werden. Eine populäre Beschleunigungstechnik für das Finden von Kürzesten Wegen ist die sogenannte Contraction Hierarchy, welche iterativ Knoten kontrahiert durch Entfernen dieser und anschließendem Einfügen von shortcuts in den Graphen. Wenn die Berechnung einer kompletten contraction hierarchy nicht möglich oder zu aufwendig ist, kann man den Graphen auch nur teilweise kontrahieren. In dieser Arbeit untersuchen wir die Leistung von Partial Contraction Hierarchies. Wir evaluieren Möglichkeiten, den Tradeoff von Speicherplatz und Beschleunigung bei contraction hierarchies zu verändern. Dies versuchen wir einerseits dadurch, Knoten unkontrahiert zu lassen und andererseits durch Entfernen von shortcuts aus dem Graph. Diesen Kompromiss haben wir systematisch untersucht und herausgefunden, dass man in einer contraction hierarchy fast nichts weglassen kann, ohne gleich viel Laufzeit zu verlieren.

Transmission Network Expansion Planning using the Railway Network

  • Date: Fri, 22 Jun 2018, 14:00
  • Location: SR 301
  • Speaker: Chao Wang, ITI Wagner; Masters Thesis
  • Inhalt: The main goal of Transmission Network Expansion Planning (TNEP) is to determine optimal investments on transmission line additions from the candidate network to satisfy reliability criteria under future load and generation scenarios. Up to now, most of the current researches focus on using different methods to find optimal solution of TNEP problem, however, how to generate the candidate network, where the new lines can be installed, has also drawn our attentions. Hence, this thesis explores a new design problem about using railway network to expand the power grid, which is defined as follows. Given the current power grid and the railway network, we build connections between both network and find, where and how many new lines along new connections and railway edges should be added with an object of minimizing the cost of lines construction under the electrical constraints. In order to build connections between the power grid and railway network, the algorithm KD-tree is proposed in our work, which can efficiently find k-nearest neighbors in railway network for each node in the power grid. For finding an optimal solution of TNEP problem, we have adopted mathematical models, such as mixed-integer non-linear model, binary linear model and mixed-integer linear model, which are been discussed in the most researches. Additionally, we have also proposed a heuristic algorithm and compared with the mathematical model. The heuristic algorithm performs more computational efficiency than the mathematical model. In the end, a simple case study of 8-buses system is used to illustrate the non-linear and linear mathematical model. Furthermore, the realistic scenarios of German including the power grid and the railway network is implemented and tested by using the linear mathematical model and the heuristic algorithm. The non-linear mathematic model comparing to the linear model is easy to understand, but has lower computational efficiency and less optimal result. On the other hand, the heuristic model shows the highest efficiency but the result is way below the ones from linear mathematic model.

Parallelizing Graphplan

  • Date: Wed, 30 May 2018, 13:00
  • Location: SR 301
  • Speaker: Patrick Hegemann, ITI Sanders; Bachelors Thesis
  • Inhalt: Automated Planning is a method used in artificial intelligence systems such as autonomous robots or automatic satellite control. While there already are some planning algorithms that make use of multi-core processors, this thesis proposes a parallelized version of the Graphplan algorithm. The proposed algorithm works by transforming plan extraction problems for different horizon lengths into the Boolean Satisfiability Problem (SAT) and solving multiple SAT formulas in parallel. Experiments based on the problem instances provided by the 2014 International Planning Competition show that this method results in a significant parallel speedup and is able to outperform the state-of-the-art SAT-based sequential planner Madagascar in some problem domains. A disadvantage of the proposed algorithm is high memory consumption for some problems with certain configurations.

Faster Public Transit Routing with Unrestricted Walking

  • Date: Wed, 23 May 2018, 13:00
  • Location: SR 301
  • Speaker: Jonas Sauer, ITI Wagner; Masters Thesis
  • Inhalt: Viele Routing-Algorithmen für öffentliche Verkehrsnetzwerke beschränken die Länge von Fußwegen, die zwischen Stationen erlaubt sind. Es wird oft argumentiert, dass lange Fußwege nur selten die Reisezeit verbessern und deswegen nicht repräsentiert werden muss. Diese Annahme wurde jedoch kürzlich in Frage gestellt. Auf der anderen Seite verwenden existierende Algorithmen, die unbeschränktes Laufen erlauben, teure Dijkstra-Suchen, um die Fußwege zu erkunden, und sind deshalb deutlich langsamer als Algorithmen mit beschränktem Laufen. Diese Arbeit beschäftigt sich deshalb damit, schnellere Routing-Algorithmen für öffentliche Verkehrsnetzwerke mit unbeschränktem Laufen zu entwickeln. Hierzu präsentieren wir eine Beschleunigungstechnik für den RAPTOR-Algorithmus, die benötigte Fußwege vorberechnet. Dadurch werden die Dijkstra-Suchen überflüssig und die Erkundung der Fußwege wird beschleunigt. Weiterhin stellen wir zwei Algorithmen vor, die an existierende Beschleunigungstechniken angelehnt sind, die für andere Szenarien entwickelt wurden: Route-Flags basiert auf Arc-Flags, während ML-RAPTOR Ideen von Customizable Route Planning und Connection Scan Accelerated aufgreift. Eine experimentelle Studie auf dem öffentlichen Verkehrsnetzwerk der Schweiz zeigt, dass die Vorberechnungstechnik für Fußwege eine deutliche Beschleunigung gegenüber existierenden Algorithmen aufweist, während Route-Flags and ML-RAPTOR nur eine geringe Beschleunigung erzielen.

On Aesthetic Functions in Visualizations of Text Variant Graphs

  • Date: Fri, 4 May 2018, 14:30
  • Location: SR 301
  • Speaker: Jonas Haas, ITI Wagner; Bachelors Thesis
  • Inhalt: Text variant graphs can assist the visualization of the differences and similarities between various texts. As a viewer is reading the original texts over the course of a text variant graph, it is important that readability of the paths given by the original texts is high. This thesis suggests aesthetic functions to improve this kind of readability in drawings of text variant graphs. Integer linear programs and a heuristic approach for implementing the metrics are provided, and the results of the optimization approaches are evaluated.

Load-Balancing for parallel Delaunay Triangulations

  • Date: Fri, 4 May 2018, 14:00
  • Location: SR 301
  • Speaker: Vincent Winkler, ITI Sanders; Bachelors Thesis
  • Inhalt: Calculating the Delaunay triangulation of large point clouds with over a million points is time-consuming. Point clouds of over a billion points may even require distributed memory. Previous work by Funke and Sanders ["Parallel d-D Delaunay Triangulations in Shared and Distributed Memory"] presents a novel parallel algorithm that calculates the Delaunay triangulation on shared or distributed memory. With more realistic inputs however, the algorithm has to reprocess about 5% of the triangulation due to its simple work division strategy. This bachelor's thesis addresses the issue to decrease the overhead and further speed up the triangulation algorithm. An elaborate load balancing approach that partitions the input point cloud in an additional preprocessing step is provided. To find the partitioning, the Delaunay triangulation of a small input sample is partitioned by a graph partitioning algorithm and the resulting partitioning is extended to a partitioning of the entire input. Different extension strategies are developed and examined for the overall runtime, the reprocessing overhead and the balance of the partition sizes. The number of multiply processed points can be reduced to less than 0.5%. Partition sizes that deviate less than 1% from each other can be achieved.

Beobachtungen des Maximalen Flexiblen Drehstromübertragungssystem Flusses

  • Date: Wed, 2 May 2018, 13:00
  • Location: SR 301
  • Speaker: Larisa Duczek, ITI Wagner; Diplomarbeit
  • Inhalt: Diese Diplomarbeit beschäftigt sich mit der Auswirkung von flexiblen Drehstromübertragungssystemen (FACTS) auf den maximalen Leistungsfluss in speziellen Netzwerken, den Kaktusnetzwerken. FACTS-Geräte erhöhen die Impedanz (Wechselstromwiderstand) auf Stromleitungen und verändern somit den Leistungsfluss. Das ermöglicht eine Erhöhung des maximalen Leistungsflusses in einem Stromnetz. Ein Kaktusnetzwerk ist ein Netzwerk, dessen Leitungen Teil höchstens eines Kreises sind. In dieser Diplomarbeit wird ein lineares Gleichstrommodell, eine Approximation des Wechselstrommodells, als Berechnungsgrundlage benutzt. Es wird der Unterschied zwischen FACTS und dem maximalen Übertragungsschaltungsfluss erläutert. Es konnte beobachtet werden, dass in manchen Netzwerken eine Mischung aus FACTS und Übertragungsschaltungen sinnvoll ist. Zudem wird gezeigt, wie groß die Auswirkungen eines FACTS auf speziellen Kreisen mit einem Erzeuger und einem Verbraucher sind.

Aesthetic Value of Graph Layouts: Investigation of Statistical Syndromes for Automatic Quantification

  • Date: Tue, 24 Apr 2018, 13:00
  • Location: SR 301
  • Speaker: Moritz Klammler, ITI Wagner; Masters Thesis
  • Inhalt: Visualizing relational data as drawing of graphs is a technique in very wide-spread use across many fields and professions. While many graph drawing algorithms have been proposed to automatically generate a supposedly high-quality picture from an abstract mathematical data structure, the graph drawing community is still searching for a way to quantify the aesthetic value of any given solution in a way that allows one to compare graph layouts created by different algorithms for the same graph (presumably to automatically choose the better one). We believe that one promising path towards this goal could be enabled by combining data analysis techniques that have proven useful in other scientific disciplines that are dealing with large structures such as astronomy, crystallography or thermodynamics. In this work we present an initial investigation of some statistical properties of graph layouts that we believe could provide viable syndromes for the aesthetic value. As a proof of concept, we used machine learning techniques to train a neural network with the results of our data analysis and thereby built a model that is able to discriminate between better and worse layouts with an accuracy of 95 %. A rudimentary evaluation of the model was performed and is presented. This work primarily provides an infrastructure to enable further experimentation on the topic and will be made available to the public as Free Software.