RESEARCH SEMINAR

A synchronous flow shop is a variant of a non-preemptive permutation flow shop where transfers of jobs from one machine to the next take place at the same time. The processing is organized in synchronized cycles which means that in a cycle all current jobs start at the same time on the corresponding machines. Then all jobs are processed and have to wait until the last one is finished. Afterwards, all jobs are moved to the next machine simultaneously. As a consequence, the processing time of a cycle is determined by the maximum processing time of the operations contained in it. Furthermore, only permutation schedules are feasible, i.e., the jobs have to be processed in the same order on all machines. The goal is to find a permutation of the jobs such that the makespan is minimized. Motivated by a practical application, we investigate special cases where the processing times of the cycles are determined by a subset of so-called dominating machines. Furthermore, we consider problems with additional resources and setup times.

With the increasing complexity of modern computing platforms, many scientific applications now rely on runtime schedulers. Most of these frameworks model an application as a directed acyclic graph of tasks, where nodes represent tasks and edges represent dependences between tasks. In my thesis, I focused on three theoretical challenges concerning task graph scheduling: exploiting the inherent parallelism of tasks, efficiently using accelerators, and coping with a limited memory. In this talk, I will detail two projects lead in this context. The first one targets a platform composed of two types of processors, such as CPUs and GPUs. The processing time of each task depends on the type of processors it is allocated to; this decision being one the main scheduling choices. We study the makespan minimization problem in an online setting, where the graph is gradually discovered, which increases the difficulty of the decision. In a second project, we focus on reducing the memory consumption of an application in order to avoid expensive memory transfers. Specifically, we aim at preventing dynamic schedulers from running out of memory on a shared-memory platform. These schedulers generally start multiple tasks in parallel, which may lead to a memory shortage. We study how to modify the task graph in order to restrict this parallelism while maintaining performance.

The Bipartite Matching Problem is a well-studied combinatorial optimization problem with numerous practical applications. In real-life optimization problems are modeled using data based on forecasts and estimates which leads to uncertain problem parameters. One way to take this fact into account is provided by the framework of robust optimization. While most robust variants of the Bipartite Matching Problem considered in the literature address cost uncertainty, this talk deals with a new robust problem, where the structure of the underlying bipartite graph is uncertain. In our problem, an additional part of the input is a list of failure scenarios described by a subset of edges of the graph that may fail simultaneously. If a scenario emerges, the corresponding edges are deleted from the graph. In the robust problem the goal is to find a minimum-cost set of edges containing a perfect matching in every scenario, i.e., a robust solution survives any possible incident. We will discuss the computational complexity and present asymptotically tight approximation algorithms. Additionally we address the following related question: Is it easier to harden an existent infrastructure rather than designing a new one from scratch? In our setting this results in the question how to robustify a given perfect matching by adding new edges to the ambient graph. For this problem we will present hardness and algorithmic results.

We address a fundamental stochastic scheduling problem where jobs with uncertain processing times must be scheduled non-preemptively on \(m\) identical parallel machines. We are given a set \(J\) of \(n\) jobs, where each job \(j\in J\) is modeled by a random variable \(P_j\) with known distribution. The actual realization of a processing time becomes known only by executing a job. The goal is to find a policy \(\Pi\) that decides for any point in time which jobs to schedule such as to minimize the expected total completion time, \(\sum_{j\in J} \mathbb{E}[C_j^\Pi]\), where \(C_j\Pi\) denotes the completion time of job \(j\) under policy \(\Pi\). The scheduling problem can be stated in the standard three-field notation as \(P || \mathbb{E}[\sum_j C_j]\). We consider so-called index policies that assign the same priority to jobs with the same probability distribution and schedule jobs one after the other on the first machine that becomes available. Job-based index policies do not consider the number of jobs or the number of machines. We rule out distribution-independent approximation factors for job-based index policies. Somewhat surprisingly this lower bound is obtained already for very simple instances with only two types of jobs, identical deterministic jobs and a set of stochastic jobs that all follow the same Bernoulli distribution. For this class of instances we also give a policy that is an \(\mathcal{O}(m)\)-approximation. This talk is based on work together with Felix Fischer, Nicole Megow, and Jannik Matuschke.

For most clustering problems, the quality of a solution is usually assessed with respect to a given pairwise distance on the input objects. Approximate solutions for such tasks often rely on this distance to be a metric. But what happens if this property does not hold? Especially for computing a clustering such that each cluster has a certain minimum cardinality k>2 (as required for anonymisation or balanced facility location problems) this becomes problematic. For example, with the objectives to minimise the maximum radius or diameter of all clusters, there exists no polynomial-time constant factor approximation if the triangle inequality is violated while there exists a 2-approximation for metric distances. This talk introduces methods to resolve or at least soften this effect of non-metric distances by devising particular strategies to deal with violations of the triangle inequality (conflicts). With the use of parameterised algorithmics it turns out that, if the number of such conflicts is not too large, constant factor approximations can still be computed efficiently.

TSP with neighborhoods is a long-studied problem where we are asked to find the shortest tour visiting a given set of regions (=neighborhoods). When the neighborhoods are lines in two dimensions, the problem is efficiently solvable. For lines in three dimensions, the problem is NP-hard and this is generally true in higher dimensions with regions that are low dimensional subspaces. The question we try to answer in this talk concerns the setting where the regions are hyperplanes in three or more dimensions. The complexity of this problem is open (no hardness results are known) and, until now, only a recent \((2^d)\)-approximation algorithm was known. In this talk, we make a step forward in answering this sixteen year old open question by designing a PTAS. In a high-level view, we will see the underlying ideas, including a (new) sparsification technique for polytopes and a linear program that computes the shortest tour.
This talk is based on work together with Antonios Antoniadis, Krzysztof Fleszar, and Kevin Schewior.

In the Survivable Network Design Problem (SNDP), we are given an undirected graph \(G\) together with costs on edges or vertices and the demand function \(k\) on source-sink demand pairs \((s_i, t_i, k_i)\), where each \({s_i, t_i}\) are vertices in the graph and \(k_i\) is a natural number. In the edge-connectivity setting (EC-SNDP), we are interested in the minimum cost subgraph \(H\) of \(G\) in which every demand pair \((s_i, t_i)\) is connected via \(k_i\) edge-disjoint paths. These problems are natural generalizations of Steiner tree and Steiner forest, which capture the fault-tolerant requirements in networks.
In general, these problems are NP-hard, and even restrictive special cases of SNDP post non-trivial challenges. As such, SNDP has drawn a lot of attention from researchers for the past decades. Special cases of SNDP, like Steiner Tree, Steiner Forest, \(2\)-Edge connected subgraph etc. has been studied extensively on many special classes of graphs, e.g. planar graphs, Euclidean metrics, and bounded treewidth graphs.
In this talk, we shall discuss some recent results on exact algorithms for the general SNDP parameterized by the treewidth of the graph. In particular, we shall demonstrate a fairly general dynamic programming framework on the basis of the tree-decomposition of the graph which can be applied to solve a bunch of related network design problems.

Modal logic and its syntactic variants are widely used in knowledge representation and verification. Applications in those areas require modal logics with a careful balance of the rivalling requirements "high expressivity" and "low computational complexity". The talk will give an overview of our contributions to "taming" expressive temporal, description, and hybrid logics, which consist in identifying fragments with good computational complexity and studying modularity. These contributions fall into three groups: (1) systematic studies of the complexity of sub-Boolean fragments of expressive modal and description logics, (2) theoretic and practical studies of module extraction and modularisation of ontologies, and (3) a study of the complexity of lightweight combinations of temporal and description logics.
The talk will be accessible to a general audience with a background in (theoretical) computer science.

New challenges need to be faced during modeling, design, implementation, verification and test of Cyber‐Physical Systems (CPS). For the design of hardware and software in this context, new approaches need to be developed, taking into account also non‐functional requirements. This talk gives an overview on challenges and recent developments with a special focus on verification and correctness.

We study the problem of finding a tour of \(n\) points in which every edge is long. More precisely, we wish to find a tour that visits every point exactly once, maximizing the length of the shortest edge in the tour. The problem is known as Maximum Scatter TSP, and was introduced by Arkin et al. (SODA 1997), motivated by applications in manufacturing and medical imaging. Arkin et al. gave a \(0.5\)-approximation for the metric version of the problem and showed that this is the best possible ratio achievable in polynomial time (assuming \(P \neq NP\)). Arkin et al. raised the question of whether a better approximation ratio can be obtained in the Euclidean plane. We answer this question in the affirmative in a more general setting, by giving a \((1-\varepsilon)\)-approximation algorithm for \(d\)-dimensional doubling metrics, with running time \( \tilde{O}(n^3 + 2^{O(K \log (K))}) \), where \(K \leq (13/\varepsilon)^d\). As a corollary we obtain (i) an efficient polynomial-time approximation scheme (EPTAS) for all constant dimensions \(d\), (ii) a polynomial-time approximation scheme (PTAS) for dimension \(d = (\log \log (n))/c\), for a sufficiently large constant \(c\), and (iii) a PTAS for constant \(d\) and \(\varepsilon = \Omega(1/\log \log (n))\). Furthermore, we show the dependence on \(d\) in our approximation scheme to be essentially optimal, unless Satisfiability can be solved in subexponential time.

In this talk, we discuss the following setting. A gambler is presented with a set of random variables that come to him in a random order. When the variable arrives, the gambler sees its value and decides to take it or go on to the next. If the gambler takes it, he gets an amount of money equal to the value of the variable and the game ends, otherwise he is presented the next random variable. The gamblers aim is to maximize the expected amount of money he receives. This setting is a classic example of an optimal stopping problem and is known as the Prophet Secretary Problem. For this problem we analyze two different threshold strategies. One non-adaptive strategy, where the thresholds are chosen before any of the variables arrive and before the order is known, and one adaptive strategy, where thresholds are chosen based on the variables that have been rejected prior. For the non-adaptive strategy we show that the gambler can, in expectation, get at least \(1-1/e\) times the expected maximum value among the random variables (i.e. the value a prophet would receive). We show that no non-adaptive strategy can do better, even for i.i.d. random variable. For the adaptive strategy, we analyze only the i.i.d. setting and show that this improves the gamblers expectation to \(0.745\) times the expected maximum. This results settles a conjecture by Kertz from 1986. Finally, we discuss the implications for posted price mechanisms, a commonly used mechanism for selling goods. We show that the results for threshold strategies imply similar bounds for comparisons of posted price mechanisms with optimal auctions.

Motivated by higher-dimensional coboundary expansion we define a certain family of graphs which are called Cheeger graphs. Little is known at present about these graphs. We will make a short self-contained introduction into the subject and formulate several open purely graph-theoretic problems.

We investigate online convex optimization with switching costs (OCO; Lin et al., INFOCOM 2011), a natural online problem arising when rightsizing data centers: A server initially located at \(p_0\) on the real line is presented with an online sequence of non-negative convex functions \(f_1,f_2,\ldots ,f_n: R\rightarrow R^+\). In response to each function \(f_i\), the server moves to a new position \(p_i\) on the real line, resulting in cost \(|p_i-p_{i-1}| + (f_ip_i)\). The total cost is the sum of costs of all steps. One is interested in designing competitive algorithms.
In this talk, we solve the problem in the classical sense: We give a lower bound of \(2\) on the competitive ratio of any possibly randomized online algorithm, matching the competitive ratio of previously known deterministic online algorithms (Andrew et al., COLT 2013/arXiv 2015; Bansal et al., APPROX 2015). It has been previously conjectured that \((2-\varepsilon)\)-competitive algorithms for some \(\varepsilon > 0\) (Bansal et al., APPROX 2015).
Joint work with Kevin Schewior, to appear in WAOA 2017.

We consider \(n\) jobs that have to be scheduled on \(m\) processors. The processors share a common, continuously divisible resource. In each discrete time step, the scheduler can change the assignment of the resource to the processors. Jobs come with a processing volume and a resource requirement. When a processor currently has a share \(R\) of the resource and works on a job with processing volume \(p\) and resource requirement \(r\), exactly \(\min\{R/r, 1\}\) units of the job's processing volume are processed. That is, doubling the resource also doubles the speed a job is processed with, but this speedup is capped. Migration and preemption are not allowed, and the goal is to minimize the makespan.
In this talk, I present an \(1 + O(1/m)\)-approximation algorithm if all jobs have unit work and an \(2 + O(1/m)\)-approximation algorithm for jobs with arbitrary workload. The algorithms employ a sliding window technique and are quite simple and intuitive. This problem generalizes bin packing with cardinality constraints and splittable items, for which our results improve the approximation ratio achieved by a fast and simple algorithm from previously \(2 - O(1/m)\) to \(1 + O(1/m)\).
This is joint work with Alexander Mäcker, Sören Riechers, and Alexander Skopalik.

We consider the problem of exploring an undirected and initially unknown graph by a set of cooperative agents. The vertices of the graph are unlabeled, and the edges incident to a vertex have locally distinct labels.In this setting, it is known that for one agent Θ(log n) bits of memory are necessary and sufficient to explore any graph with at most n vertices. We show that O(log log n) cooperative agents with only constant memory can explore any graph with at most n vertices in polynomial time. We match this result with a lower bound showing that the number of agents needed for exploring any graph with at most n vertices is already Ω(log log n) when we allow each agent to have at most O(log n 1−ε ) bits of memory for some ε > 0. In particular, this implies that Ω(log log n) agents with constant memory are necessary for exploration.

Submodular functions are set functions that capture the natural "diminishing returns" property and can be seen as a discrete analogy to convex functions. The problem of maximizing a submodular function is as follows: Given a set \(E\) of elements and a set function \(f: 2^E\rightarrow Q\) (typically through an oracle), we want to determine a subset \(S\) of \(E\) with maximum value \(f(S)\). This is an NP-hard problem for which several approximation algorithms are known. These algorithms are either randomized or adaptive, in the sense that they choose which function values to look up after having already seen some others, and in this way adapt to the function. In this talk we discuss a non-adaptive, or oblivious, deterministic approach. All sets for which the value of the function will be queried will be chosen from the beginning as a list. The set of this list with the highest function value will then be returned as an approximation of the problem. The focus of the talk will be choosing such a list and determining the quality of this solution.

This will be a talk covering my general research interests, approximation and online algorithms for scheduling and network design.

First, I will speak about online scheduling problems. We propose a new model on top of the classical competitive analysis that we believe is useful in dealing with uncertainty in the input data. We show that one can design simple online algorithms with good guarantees if one is allowed to reject, that is, not process a tiny fraction of the input. We propose this in the light of the fact that such an algorithm is not possible for optimizing certain classical scheduling objectives like makespan minimization or minimizing the (weighted) maximum flow-time.

Joint work with: Anamitra Roy Choudhury, Naveen Garg and Amit Kumar.

In the second part, I shall talk about network design problems. We give improved approximation guarantees for a classical network design problem called the Group Steiner Tree, when the input graph has a special structure. In particular, we prove a generic reduction of a problem instance on such graphs to a problem instance on trees that allows us to carry out the desired improvement. The reduction may be of independent interest.

Joint work with: Parinya Chalermsook, Bundit Laekhanukit and Daniel Vaz.

Minimizing resource usage is a key to achieving economic, environmental, or societal goals. We consider the fundamental problem of minimizing the number of machines that is necessary for feasibly scheduling preemptive jobs with release dates and hard deadlines. We consider the online variant of this problem in which every job becomes known to the online algorithm only at its release date. We present new competitive algorithms for this problem. We also discuss the power of job migration and show somewhat surprising strong lower bounds on the best possible performance guarantee.

Joint work with: Lin Chen and Kevin Schewior.

Consider the following variant of the set cover problem. We are given a universe \(U=\{1,...,n\}\) and a collection of subsets \(\mathcal{C} = \{S_1,...,S_m\}\) where \(S_i \subseteq U\). For every element \(u \in U\) we need to find a set \(\phi(u) \in \mathcal C\) such that \(u\in \phi(u)\). Once we construct and fix the mapping \(\phi:U \mapsto \mathcal C\) a subset \(X \subseteq U\) of the universe is revealed, and we need to cover all elements from \(X\) with exactly \(\phi(X):=\bigcup_{u\in X} \phi(u)\). The goal is to find a mapping such that the cover \(\phi(X)\) is as cheap as possible. This is an example of a universal problem where the solution has to be created before the actual instance to deal with is revealed. Such problems appear naturally in some settings when we need to optimize under uncertainty and it may be actually too expensive to begin finding a good solution once the input starts being revealed. A rich body of work was devoted to investigate such problems under the regime of worst case analysis, i.e., when we measure how good the solution is by looking at the worst-case ratio: universal solution for a given instance vs optimum solution for the same instance. As the universal solution is significantly more constrained, it is typical that such a worst-case ratio is actually quite big. One way to give a viewpoint on the problem that would be less vulnerable to such extreme worst-cases is to assume that the instance, for which we will have to create a solution, will be drawn randomly from some probability distribution. In this case one wants to minimize the expected value of the ratio: universal solution vs optimum solution. Here the bounds obtained are indeed smaller than when we compare to the worst-case ratio. But even in this case we still compare apples to oranges as no universal solution is able to construct the optimum solution for every possible instance. What if we would compare our approximate universal solution against an optimal universal solution that obeys the same rules as we do? In the talk we will see that under this viewpoint, but still in the stochastic variant, we can indeed obtain better bounds than in the expected ratio model. For the Set Cover problem --- on which the main ideas and characteristics of the model will be presented --- we will see how one can obtain \(H_n\) approximation, which matches the approximation ratio from the classic deterministic setup. Moreover, this result holds for all possible probability distributions over \(U\) that have a polynomially large carrier, while all previous results pertained to a model in which elements were sampled independently. The result is based on rounding a proper configuration IP that captures the optimal universal solution, and using tools from submodular optimization.

Joint work with: Fabrizio Grandoni, Stefano Leonardi and Michał Włodarczyk.

The area of mechanism design is concerned with the development of algorithms that produce good outcomes given inputs from self-interested individuals. Much of mechanism design theory is dominated by the Vickrey-Clarke-Groves (VCG) mechanism, which makes it optimal for each individual to provide its true input. In the real world the VCG mechanism is used with surprising rarity, and some reasons for this are known. We identify another property of the mechanism that may be problematic in practice, a relative lack of robustness to inaccuracies in the choice of its parameters.

Joint work with Paul Dütting and David C. Parkes.