Search

Security has become a critical issue for complex and expensive systems and day-to-day situations. In this regard, the analysis of surveillance cameras is a critical issue usually restricted to the number of people devoted to such a task, their knowledge and judgment. Nonetheless, different approaches have arisen to automate this task in recent years. These approaches are mainly based on machine learning and benefit from developing neural networks capable of extracting underlying information from input videos. Despite how competent those networks have proved to be, developers must face the challenging task of defining both the architecture and hyperparameters that allow such networks to work adequately and optimize the use of computational resources. In short, this work proposes a model that generates, through a genetic algorithm, neural networks for behavior classification within videos. Two types of neural networks evolved as part of this work, shallow and deep, which are structured on dense and 3D convolutional layers. Each network requires a particular type of input data: the evolution of the pose of people in the video and video sequences, respectively. Shallow neural networks use a direct encoding approach to map each part of the chromosome into a phenotype. In contrast, deep neural networks use indirect encoding, blueprints representing entire networks, and modules to depict layers and their connections. Our approach obtained relevant results when tested on the Kranok-NV dataset and evaluated with standard metrics used for similar classification tasks. [ABSTRACT FROM AUTHOR]

Metaheuristics have become a widely used approach for solving a variety of practical problems. The literature is full of diverse metaheuristics based on outstanding ideas and with proven excellent capabilities. Nonetheless, oftentimes metaheuristics claim novelty when they are just recombining elements from other methods. Hence, the need for a standard metaheuristic model is vital to stop the current frenetic tendency of proposing methods chiefly based on their inspirational source. This work introduces a first step to a generalised and mathematically formal metaheuristic model, which can be used for studying and improving them. This model is based on a scheme of simple heuristics, which perform as building blocks that can be modified depending on the application. For this purpose, we define and detail all components and concepts of a metaheuristic (i.e., its search operators), such as heuristics. Furthermore, we also provide some ideas to take into account for exploring other search operator configurations in the future. To illustrate the proposed model, we analyse search operators from four well-known metaheuristics employed in continuous optimisation problems as a proof-of-concept. From them, we derive 20 different approaches and use them for solving some benchmark functions with different landscapes. Data show the remarkable capability of our methodology for building metaheuristics and detecting which operator to choose depending on the problem to solve. Moreover, we outline and discuss several future extensions of this model to various problem and solver domains. [ABSTRACT FROM AUTHOR]

Highlights • A new method to solve vehicle routing through meta-learning techniques. • Two sets of meta-features are used: basic and landmarking meta-features. • A multilayer perceptron classifier is used to select meta-heuristics. • Our proposal statistically improves the overall performances previously reported. Abstract This paper describes a method for solving vehicle routing problems with time windows, based on selecting meta-heuristics via meta-learning. Although several meta-heuristics already exist that can obtain good overall results on some vehicle routing problem instances, none of them performs well in all cases. By defining a set of meta-features that appropriately characterize different routing problem instances and using a suitable classifier, our model can often correctly predict the best meta-heuristic for each instance. The main contributions of this paper are the definition of two meta-feature sets, one based on what we call 'basic' instance properties and another based on the number of feasible solutions found by perturbative heuristics via a greedy process. We use a multilayer perceptron classifier, combined with a wrapper meta-feature selection method, to predict the most suitable meta-heuristic to apply to a given problem instance. Our experimental results show that the proposed method can significantly improve upon the overall performance of well-known meta-heuristics in the field. Therefore, this paper proposes to store, share and exploit in an off-line scheme the solutions obtained in instances of different scenarios such as the academy or industry, with the aim of predicting good solvers for new instances when necessary. [ABSTRACT FROM AUTHOR]

• Features of initial solutions predict the performance of perturbation algorithms. • Low performance variation within portfolios makes algorithm selection difficult. • Hybrid selection models are useful for portfolios with low performance variation. • Accuracy is not a fair measure to evaluate the performance of algorithm selectors. In this paper, we present the construction process of a per-instance algorithm selection model to improve the initial solutions of Curriculum-Based Course Timetabling (CB-CTT) instances. Following the meta-learning framework, we apply a hybrid approach that integrates the predictions of a classifier and linear regression models to estimate and compare the performance of four meta-heuristics across different problem sub-spaces described by seven types of features. Rather than reporting the average accuracy, we evaluate the model using the closed SBS-VBS gap, a performance measure used at international algorithm selection competitions. The experimental results show that our model obtains a performance of 0.386, within the range obtained by per-instance algorithm selection models in other combinatorial problems. As a result of the process, we conclude that the performance variation between the meta-heuristics has a significant role in the effectiveness of the model. Therefore, we introduce statistical analyses to evaluate this factor within per-instance algorithm portfolios. [ABSTRACT FROM AUTHOR]