International Journal on Artificial Intelligence Tools

We study the problem of finding the global optimum of a nonlinear real function over an interval box by means of complete search techniques, namely interval branch-and-bound algorithms. Such an algorithm typically generates a tree of boxes from the initial box by alternating branching steps and contraction steps in order to remove non optimal sub-boxes. In this paper, we introduce a new contraction method that is designed to handle the boundary of the initial box where a minimizer may not be a stationary point. This method exploits the first-order optimality conditions and we show that it subsumes the classical monotonicity test based on interval arithmetic. A new branch-and-bound algorithm has been implemented in the interval solver Realpaver. An extensive experimental study based on a set of standard benchmarks is presented.

Deep neural networks become more popular as its ability to solve very complex pattern recognition problems. However, deep neural networks often need massive computational and memory resources, which is main reason resulting them to be difficult efficiently and entirely running on embedded platforms. This work addresses this problem by saving the computational and memory requirements of deep neural networks by proposing a variance reduced (VR)-based optimization with regularization techniques to compress the requirements of memory of models within fast training process. It is shown theoretically and experimentally that sparsity-inducing regularization can be effectively worked with the VR-based optimization whereby in the optimizer the behaviors of the stochastic element is controlled by a hyper-parameter to solve non-convex problems.

We propose a logical encoding of argumentation frameworks with higher-order interactions (i.e. attacks/supports whose targets are arguments or other attacks/supports) with an evidential meaning for supports. Our purpose is to separate the logical expression of the meaning of an attack or an evidential support (simple or higher-order) from the logical expression of acceptability semantics. We consider semantics which specify the conditions under which the arguments (resp. the attacks/supports) are considered as accepted, directly on the extended framework, without translating the original framework into a Dung’s argumentation framework. We characterize the output of a given framework in logical terms (namely as particular models of a logical theory). Our proposal applies to the particular case of Dung’s frameworks, enabling to recover standard extensions.

Centrality measures have been helping to explain the behavior of objects, given their relation, in a wide variety of problems, since sociology to chemistry. This work considers these measures to assess the importance of every classifier belonging to an ensemble of classifiers, aiming to improve a Multiple Classifier System (MCS). Assessing the classifier’s importance by employing centrality measures, inspired two different approaches: one for selecting classifiers and another for fusion. The selection approach, called Centrality Based Selection (CBS), adopts a trade-off between the classifier’s accuracy and their diversity. The sub-optimal selected subset presents good results against selection methods from the literature, being superior in 67.22% of the cases. The second approach, the integration, is named Centrality Based Fusion (CBF). This approach is a weighted combination method, which is superior to literature in 70% of the cases.

Possibilistic graphical models are powerful modeling and reasoning tools for uncertain information based on possibility theory. In this paper, we provide an analysis of computational complexity of MAP and MPE queries for possibilistic networks. MAP queries stand for maximum a posteriori explanation while MPE ones stand for most plausible explanation. We show that the decision problems of answering MAP and MPE queries in both min-based and product-based possibilistic networks is NP-complete. Definitely, such results represent an advantage of possibilistic graphical models over probabilistic ones since MAP queries are NP^PP -complete in Bayesian networks. Our proofs for querying min-based possibilistic networks make use of reductions from the 3SAT problem to querying possibilistic networks decision problem. Moreover, the provided reductions may be useful for the implementation of MAP and MPE inference engines based on the satisfiability problem solvers. As for product-based networks, the provided proofs are incremental and make use of reductions from SAT and its weighted variant WMAXSAT.

Constraint Programming (CP) is a powerful declarative programming paradigm where inference and search are interleaved to find feasible and optimal solutions to various type of constraint systems. However, handling logical connectors with constructive information in CP is notoriously difficult. This paper presents If Then Else (ITE), a lightweight implementation of stratified constructive reasoning for logical connectives. Stratification is introduced to cope with the risk of combinatorial explosion of constructing information from nested and combined logical operators. ITE is an open-source library built on top of SICStus Prolog clpfd, which proposes various operators, including constructive disjunction and negation, constructive implication and conditional. These operators can be used to express global constraints and to benefit from constructive reasoning for more domain pruning during constraint filtering. Even though ITE is not competitive with specialized filtering algorithms available in some global constraints implementations, its expressiveness allows users to easily define well-tuned constraints with powerful deduction capabilities. Our extended experimental results show that ITE is more efficient than available generic approaches that handle logical constraint systems over finite domains.

This paper^a describes a maintenance scheduling application, which was developed together with an industrial partner. This is a highly combinatorial decision process, to plan and schedule the work of a group of travelling repair technicians, which perform preventive and corrective maintenance tasks at customer locations. Customers are located both in urban areas, where many customers are in close proximity, and in sparsely populated rural areas, where the travel time between customer sites is significant. To balance the workload for the agents, we must consider both the productive working time, as well as the travel between locations. As the monolithic problem formulation is unmanageable, we introduce a problem decomposition into multiple sequential steps, that is compatible with current management practice. We present and compare different models for the solution steps, and discuss results on datasets provided by the industrial partner.

Big Data analytics is presently one of the most emerging areas of research for both organizations and enterprises. The requirement for deployment of efficient machine learning algorithms over huge amounts of data led to the development of parallelization frameworks and of specialized libraries (like Mahout and MLlib) which implement the most important among these algorithms. Moreover, the recent advances in storage technology resulted in the introduction of high-performing devices, broadly known as Solid State Drives (SSDs). Compared to the traditional Hard Drives (HDDs), SSDs offer considerably higher performance and lower power consumption. Motivated by these appealing features and the growing necessity for efficient large-scale data processing, we compared the performance of several machine learning algorithms on MapReduce clusters whose nodes are equipped with HDDs, SSDs, and devices which implement the latest 3D XPoint technology. In particular, we evaluate several dataset preprocessing methods like vectorization and dimensionality reduction, two supervised classifiers, Naive Bayes and Linear Regression, and the popular k-Means clustering algorithm. We use an experimental cluster equipped with the three aforementioned storage devices under different configurations, and two large datasets, Wikipedia and HIGGS. The experiments showed that the benefits which derive from the usage of SSDs depend on the cluster setup and the nature of the applied algorithms.

Social media contain rich information that can be used to help understand human mind and behavior. Social media data, however, are mostly unstructured (e.g., text and image) and a large number of features may be needed to represent them (e.g., we may need millions of unigrams to represent social media texts). Moreover, accurately assessing human behavior is often difficult (e.g., assessing addiction may require medical diagnosis). As a result, the ground truth data needed to train a supervised human behavior model are often difficult to obtain at a large scale. To avoid overfitting, many state-of-the-art behavior models employ sophisticated unsupervised or self-supervised machine learning methods to leverage a large amount of unsupervised data for both feature learning and dimension reduction. Unfortunately, despite their high performance, these advanced machine learning models often rely on latent features that are hard to explain. Since understanding the knowledge captured in these models is important to behavior scientists and public health providers, we explore new methods to build machine learning models that are not only accurate but also interpretable. We evaluate the effectiveness of the proposed methods in predicting Substance Use Disorders (SUD). We believe the methods we proposed are general and applicable to a wide range of data-driven human trait and behavior analysis applications.

In this paper we describe a general framework for parallel optimization based on the island model of evolutionary algorithms. The framework runs a number of optimization methods in parallel with periodic communication. In this way, it essentially creates a parallel ensemble of optimization methods. At the same time, the system contains a planner that decides which of the available optimization methods should be used to solve the given optimization problem and changes the distribution of such methods during the run of the optimization. Thus, the system effectively solves the problem of online parallel portfolio selection.

The proposed system is evaluated in a number of common benchmarks with various problem encodings as well as in two real-life problems — the optimization in recommender systems and the training of neural networks for the control of electric vehicle charging.

During the last years, big data has become the new emerging trend that increasingly attracting the attention of the R&D community in several fields (e.g., image processing, database engineering, data mining, artificial intelligence). Marine data is part of these fields which accommodates this growth, hence the appearance of marine big data paradigm that monitoring advocates the assessment of human impact on marine data. Nonetheless, supporting acoustic sounds classification is missing in such environment, with taking into account the diversity of such data (i.e., sounds of living undersea species, sounds of human activities, and sounds of environmental effects). To overcome this issue, we propose in this paper an approach that efficiently allowing acoustic diversity classification using machine learning techniques. The aim is to reach an automated support of marine big data analysis. We have conducted a set of experiments, using a real marine dataset, in order to validate our approach and show its effectiveness and efficiency. To do so, three machine learning techniques are employed: (i) classic machine learning models (i.e., k-nearest neighbor and support vector machine), (ii) deep learning based on convolutional neural networks, and (iii) transfer learning based on the reuse of pretrained models.

We introduce an enhanced notion of unsatisfiable cores for QBF in prenex CNF that allows to weaken universal quantifiers to existential quantifiers in addition to the traditional removal of clauses. The resulting unsatisfiable cores can be different from those of the traditional notion in terms of syntax, standard semantics, and proof-based semantics. This not only gives rise to explanations of unsatisfiability but, via duality, also leads to diagnoses and repairs of unsatisfiability that are not obtained with traditional unsatisfiable cores. We use a source-to-source transformation on QBF in PCNF such that the weakening of universal quantifiers to existential quantifiers in the original formula corresponds to the removal of clauses in the transformed formula. This makes any tool or method for the computation of unsatisfiable cores of the traditional notion available for the computation of unsatisfiable cores of our enhanced notion. We implement our approach as an extension to the QBF solver DepQBF, and we perform an extensive experimental evaluation on a subset of QBFLIB. We illustrate with several case studies that helpful information can be provided by unsatisfiable cores of our enhanced notion.