Serdar Kadioglu

Algorithm selection and scheduling

Algorithm portfolios aim to increase the robustness of our ability to solve problems efficiently. While recently proposed algorithm selection methods come ever closer to identifying the most appropriate solver given an input instance, they are bound to make wrong and, at times, costly decisions. Solver scheduling has been proposed to boost the performance of algorithm selection. Scheduling tries to allocate time slots to the given solvers in a portfolio so as to maximize, say, the number of solved instances within a given time limit. We show how to solve the corresponding optimization problem at a low computational cost using column generation, resulting in fast and high quality solutions. We integrate this approach with a recently introduced algorithm selector, which we also extend using other techniques. We propose various static as well as dynamic scheduling strategies, and demonstrate that in comparison to pure algorithm selection, our novel combination of scheduling and solver selection can significantly boost performance.

Feature Selection Methods and Their Combinations in High-Dimensional Classification of Speaker Likability, Intelligibility and Personality Traits

This study focuses on feature selection in paralinguistic analysis and presents recently developed supervised and unsupervised methods for feature subset selection and feature ranking. Using the standard k-nearest-neighbors (kNN) rule as the classification algorithm, the feature selection methods are evaluated individually and in different combinations in seven paralinguistic speaker trait classification tasks. In each analyzed data set, the overall number of features highly exceeds the number of data points available for training and evaluation, making a well-generalizing feature selection process extremely difficult. The performance of feature sets on t feature selection data is observed to be a poor indicator of their performance on unseen data. The studied feature selection methods clearly outperform a standard greedy hill-climbing selection algorithm by being more robust against overfitting. When the selection methods are suitably combined with each other, the performance in the classification task can be further improved. In general, it is shown that the use of automatic feature selection in paralinguistic analysis can be used to reduce the overall number of features to a fraction of the original feature set size while still achieving a comparable or even better performance than baseline support vector machine or random forest classifiers using the full feature set. The most typically selected features for recognition of speaker likability, intelligibility and five personality traits are also reported.

Dichotomic Search Protocols for Constrained Optimization

We devise a theoretical model for dichotomic search algorithms for constrained optimization. We show that, within our model, a certain way of choosing the breaking point minimizes both expected as well as worst case performance in a skewed binary search. Furthermore, we show that our protocol is optimal in the expected and in the worst case. Experimental results illustrate performance gains when our protocols are used within the search strategy by Streeter and Smith.

DASH: Dynamic Approach for Switching Heuristics

Complete tree search is a highly effective method for tackling Mixed-Integer Programming (MIP) problems, and over the years, a plethora of branching heuristics have been introduced to further refine the technique for varying problems. Yet while each new approach continued to push the state-of-the-art, parallel research began to repeatedly demonstrate that there is no single method that would perform the best on all problem instances. Tackling this issue, portfolio algorithms took the process a step further, by trying to predict the best heuristic for each instance at hand. However, the motivation behind algorithm selection can be taken further still, and used to dynamically choose the most appropriate algorithm for each encountered sub-problem. In this paper we identify a feature space that captures both the evolution of the problem in the branching tree and the similarity among sub-problems of instances from the same MIP models. We show how to exploit these features on-the-fly in order to decide the best time to switch the branching variable selection heuristic and then show how such a system can be trained efficiently. Experiments on a highly heterogeneous collection of hard MIP instances show significant gains over the standard pure approach which commits to a single heuristic throughout the search.

Incorporating variance in impact-based search

We present a simple modification to the idea of impact-based search which has proven highly effective for several applications. Impacts measure the average reduction in search space due to propagation after a variable assignment has been committed. Rather than considering the mean reduction only, we consider the idea of incorporating the variance in reduction. Experimental results show that using variance can result in improved search performance.

Same-relation constraints

The ALLDIFFERENT constraint was one of the first global constraints [17] and it enforces the conjunction of one binary constraint, the not-equal constraint, for every pair of variables. By looking at the set of all pairwise not-equal relations at the same time, AllDifferent offers greater filtering power. The natural question arises whether we can generally leverage the knowledge that sets of pairs of variables all share the same relation. This paper studies exactly this question. We study in particular special constraint graphs like cliques, complete bipartite graphs, and directed acyclic graphs, whereby we always assume that the same constraint is enforced on all edges in the graph. In particular, we study whether there exists a tractable GAC propagator for these global Same-Relation constraints and show that AllDifferent is a huge exception: most Same-Relation Constraints pose NP-hard filtering problems. We present algorithms, based on AC-4 and AC-6, for one family of Same-Relation Constraints, which do not achieve GAC propagation but outperform propagating each constraint individually in both theory and practice.

Availability Optimization in Cloud-Based In-Memory Data Grids

This paper presents a Constraint Programming (CP)-based application for dynamic cache distribution in Oracle Coherence In-Memory Data Grid (IMDG). A re-sizable decomposition method using CP is developed to ensure high availability through incremental optimization of load distribution and data replication. The application highlights the flexibility and efficiency that the CP technology offers for (1) concisely capturing the multiple dynamic aspects and complex constraints of the Oracle Coherence IMDG cache distribution problem; and (2) solving large-scale problem instances in a dynamic cloud environment. Extensive computational results are presented to assess the scalability and efficiency of the proposed solution.

Heterogeneous resource allocation in Cloud Management

This paper introduces a combinatorial problem arising from real-world business requirements as part of resource allocation in Cloud Management. In particular, we focus on the allocation of a set of heterogeneous resources serving multiple tenants with different service level agreements. There exist certain business rules that govern the application stemming from privacy, performance, and capacity requirements. We show how to formulate the problem as constrained optimization and then solve it efficiently using Artificial Intelligence based constraint propagation. Our approach stands out as a high-level, declarative solution that is efficient and easy to maintain and update.

Learning a Reactive Restart Strategy to Improve Stochastic Search

Building on the recent success of bet-and-run approaches for restarted local search solvers, we introduce the idea of learning online adaptive restart strategies. Universal restart strategies deploy a fixed schedule that runs with utter disregard of the characteristics that each individual run exhibits. Whether a run looks promising or abysmal, it gets run exactly until the predetermined limit is reached. Bet-and-run strategies are at least slightly less ignorant as they decide which trial to use for a long run based on the performance achieved so far. We introduce the idea of learning fully adaptive restart strategies for black-box solvers, whereby the learning is performed by a parameter tuner. Numerical results show that adaptive strategies can be learned effectively and that these significantly outperform bet-and-run strategies.

Optimizing the cloud service experience using constraint programming

This paper shows how to model and solve an important application of the well-known assignment problem that emerges as part of workforce management, particularly in cloud based customer service center optimization. The problem consists of matching a set of highly skilled agents to a number of incoming requests with specialized requirements. The problem manifests itself in a fast-paced online setting, where the complete set of incoming requests is not known apriori, turning this into a challenging problem where rapid response time and quality of assignments are crucial for success and customer satisfaction. The key contribution of this paper lies in the combination of a high-level constraint model with customizable search that can take into account various objective criteria. The result is an efficient and flexible solution that excels in dynamic environments with complex, conflicting and often changing requirements. The constraint programming approach handles hundreds of incoming requests in real-time while ensuring high-quality agent assignments.