Annals of Mathematics and Artificial Intelligence最新文献

In this paper, we develop two SVM-based classifiers named stable nested one-class support vector machines (SN-1SVMs) and decoupled margin-moment based SVMs (DMMB-SVMs), to predict the specific type of pancreatic carcinoma using quantitative histopathological signatures of images. For each patient, the diagnosis can produce hundreds of images, which can be used to classify the pancreatic tissues into three classes: chronic pancreatitis, intraductal papillary mucinous neoplasms, and pancreatic carcinoma. The proposed two approaches tackle the classification problems from two different perspectives: the SN-1SVM treats each image as a classification point in a nested fashion to predict malignancy of the tissues, while the DMMB-SVM treats each patient as a classification point by assembling information across images. One attractive feature of the DMMB-SVM is that, in addition to utilizing the mean information, it also takes into account the covariance of features extracted from images for each patient. We conduct numerical experiments to evaluate and compare performance of the two methods. It is observed that the SN-1SVM can take advantage of the data structure more effectively, while the DMMB-SVM demonstrates better computational efficiency and classification accuracy. To further improve interpretability of the final classifier, we also consider the (ell _1)-norm in the DMMB-SVM to handle feature selection.

There is a parameter ubiquitous throughout the deep learning world: learning rate. There is likewise a ubiquitous question: what should that learning rate be? The true answer to this question is often tedious and time consuming to obtain, and a great deal of arcane knowledge has accumulated in recent years over how to pick and modify learning rates to achieve optimal training performance. Moreover, the long hours spent carefully crafting the perfect learning rate can come to nothing the moment your network architecture, optimizer, dataset, or initial conditions change ever so slightly. But it need not be this way. We propose a new answer to the great learning rate question: the Autonomous Learning Rate Controller. Find it at https://github.com/fastestimator/ARC/tree/v2.0.

Strategic spatial planning is becoming more popular around the world as a decision-making way to build a unified vision for directing the medium- to long-term development of land/marine areas. Recently, the study of marine areas in terms of spatial planning such as Marine Spatial Planning (MSP) has received much attention. One of the challenging issues in MSP is to make a balance between determining the ideal zone for a new activity while also considering the locations of existing activities. This spatial zoning problem for multi-uses with multiple objectives could be formulated as optimization models. This paper presents and compares the results of two multi-objective evolutionary-based algorithms (MOEAs), Synchronous Hypervolume-based non-dominated sorting genetic algorithm-II (SH-NSGA-II) which is an extension of NSGA-II and a memetic algorithm (MA) in which SH-NSGA-II is enhanced with a local search. These proposed algorithms are used to solve the multi-objective spatial zoning optimization problem, which seeks to maximize the zone interest value assigned to the new activity while simultaneously maximizing its spatial compactness. We introduce several innovations in these proposed algorithms to address the problem constraints and to improve the robustness of the traditional NSGA-II and MA approaches. Unlike traditional ones, a different stop condition, multiple crossover, mutation, and repairing operators, and also a local search operator are developed. A comparative study is presented between the results obtained using both algorithms. To guarantee robust results for both algorithms, their parameters are calibrated and tuned using the Multi-Response Surface Methodology (MRSM) method. The effective and non-effective components, as well as the validity of the regression models, are determined using analysis of variance (ANOVA). Although SH-NSGA-II has revealed a good efficiency, its performance is still improved using a local search scheme within SH-NSGA-II, which is specially tailored to the problem characteristics. The two methods are designed for raster data.

Many recent studies have shown that Fractal Dimension (FD), a ratio for figuring out the complexity of a system given its measurements, can be used as an useful index to provide information about certain brain disease. Our research focuses on the Alzheimer’s disease changes in white and grey brain matters detected through the FD indexes of their contours. Data used in this study were obtained from the Alzheimer’s Disease (AD) Neuroimaging Initiative database (Normal Condition, N = 57, and Alzheimer’s Disease, N = 60). After standard preprocessing pipeline, the white and grey matter 3D FD indexes are computed for the two groups. A statistical analysis shows that only grey matter 3D FD indexes are able to differentiate healthy and AD subjects. Although white matter 3D FD indexes do not, it is remarkable that their presence enhance the separation capability of previous ones. In order to valuate the classification capability of these indexes on healthy and AD subjects, we define several Neural Networks models. The performances of these models vary according to the statistical analysis and reach their best performances when each 3D FD input index is changed into a sequence of 2D FD indexes of (a subset of) the horizontal slices of the white and grey matter volumes.

Deciphering gene regulatory networks’ functioning is an essential step for better understanding of life, as these networks play a fundamental role in the control of cellular processes. Boolean networks have been widely used to represent gene regulatory networks. They allow to describe the dynamics of complex gene regulatory networks straightforwardly and efficiently. The attractors are essential in the analysis of the dynamics of a Boolean network. They explain that a particular cell can acquire specific phenotypes that may be transmitted over several generations. In this work, we consider a new representation of Boolean networks’ dynamics based on a new semantics used in Answer Set Programming (ASP). We use logic programs and ASP to express and deal with gene regulatory networks seen as Boolean networks, and develop a method to detect all the attractors of such networks. We first show how to represent and deal with general Boolean networks for the synchronous and asynchronous updates modes, where the computation of attractors requires a simulation of these networks’ dynamics. Then, we propose an approach for the particular case of circular networks where no simulation is needed. This last specific case plays an essential role in biological systems. We show several theoretical properties; in particular, simple attractors of the gene networks are represented by the stable models of the corresponding logic programs and cyclic attractors by its extra-stable models. These extra-stable models correspond to the extra-extensions of the new semantics that are not captured by the semantics of stable models. We then evaluate the proposed approach for general Boolean networks on real biological networks and the one dedicated to the case of circular networks on Boolean networks generated randomly. The obtained results for both approaches are encouraging.

Nguyen et al. (2016) introduced altruistic hedonic games in which agents’ utilities depend not only on their own preferences but also on those of their friends in the same coalition. We propose to extend their model to coalition formation games in general, considering also the friends in other coalitions. Comparing our model to altruistic hedonic games, we argue that excluding some friends from the altruistic behavior of an agent is a major disadvantage that comes with the restriction to hedonic games. After introducing our model and showing some desirable properties, we additionally study some common stability notions and provide a computational analysis of the associated verification and existence problems.

In recent years, multi-task learning (MTL) has become a popular field in machine learning and has a key role in various domains. Sharing knowledge across tasks in MTL can improve the performance of learning algorithms and enhance their generalization capability. A new approach called the multi-task least squares twin support vector machine (MTLS-TSVM) was recently proposed as a least squares variant of the direct multi-task twin support vector machine (DMTSVM). Unlike DMTSVM, which solves two quadratic programming problems, MTLS-TSVM solves two linear systems of equations, resulting in a reduced computational time. In this paper, we propose an enhanced version of MTLS-TSVM called the improved multi-task least squares twin support vector machine (IMTLS-TSVM). IMTLS-TSVM offers a significant advantage over MTLS-TSVM by operating based on the empirical risk minimization principle, which allows for better generalization performance. The model achieves this by including regularization terms in its objective function, which helps control the model’s complexity and prevent overfitting. We demonstrate the effectiveness of IMTLS-TSVM by comparing it to several single-task and multi-task learning algorithms on various real-world data sets. Our results highlight the superior performance of IMTLS-TSVM in addressing multi-task learning problems.

A Digital Twin (DT) is a frequently updated virtual representation of a physical or a digital instance that captures its properties of interest. Incorporating both cyber and physical parts to build a digital twin is challenging due to the high complexity of the requirements that should be addressed and satisfied during the design, implementation and operation. In this context, we introduce the MADTwin (Multi-Agent Digital Twin) framework driven by a Multi-agent Systems (MAS) paradigm and supported by flexible architecture and extendible upper ontology for modelling agent-based digital twins. A comprehensive case study of a smart warehouse supported by multi-robots has been presented to show the feasibility and applicability of this framework. The introduced framework powered by intelligent agents integrated with enabler technologies enabled us to cope with parts of the challenges imposed by modelling and integrating Cyber-Physical Systems (CPS) with digital twins for multi-robots of the smart warehouse. In this framework, different components of CPS (robots) are represented as autonomous physical agents with their digital twin agents in the digital twin environment. Agents act autonomously and cooperatively to achieve their local goals and the objectives of the whole system. Eventually, we discuss the framework’s strengths and identify areas of improvement and plans for future work.

Digital Twins (DTs) are emerging as a fundamental brick of engineering Cyber-Physical Systems (CPSs), but their notion is still mostly bound to specific business domains (e.g. manufacturing), goals (e.g. product design), or applications (e.g. the Internet of Things). As such, their value as general purpose engineering abstractions is yet to be fully revealed. In this paper, we relate DTs with agents and multiagent systems, as the latter are arguably the most rich abstractions available for the engineering of complex socio-technical and CPSs, and the former could both fill in some gaps in agent-oriented software engineering and benefit from an agent-oriented interpretation—in a cross-fertilisation journey.

Gaussian Process based Bayesian Optimization is largely adopted for solving problems where the inputs are in Euclidean spaces. In this paper we associate the inputs to discrete probability distributions which are elements of the probability simplex. To search in the new design space, we need a distance between distributions. The optimal transport distance (aka Wasserstein distance) is chosen due to its mathematical structure and the computational strategies enabled by it. Both the GP and the acquisition function is generalized to an acquisition functional over the probability simplex. To optimize this functional two methods are proposed, one based on auto differentiation and the other based on proximal-point algorithm and the gradient flow. Finally, we report a preliminary set of computational results on a class of problems whose dimension ranges from 5 to 100. These results show that embedding the Bayesian optimization process in the probability simplex enables an effective algorithm whose performance over standard Bayesian optimization improves with the increase of problem dimensionality.