نشریه محاسبات نرم
سال سیزدهم شماره 2 (پیاپی 26، پاییز و زمستان 1403)

Modularization methods are utilized to divide program artifacts into comprehensible modules to extract software architecture. Software modularization problem is NP-hard, so search-based techniques are commonly used to address it. In recent years, the use of hyper-heuristics with intelligent search approaches has increased to achieve a higher level of generality. In this paper, a general hyper-heuristic based on the multi-agent systems is proposed for software modularization. In the proposed algorithm, agents with diversification and intensification search perspectives are used, and agents with the same perspective are placed in a coalition. In each step of the search, the most appropriate coalition is automatically selected using reinforcement learning, and its agents are executed in parallel. Also, in the design of some agents, to maintain diversity, the concept of chaos theory is used. To indicate the applicability of the suggested algorithm, we selected eleven real-world applications of small and medium sizes, along with ten folders of Mozilla Firefox that have various functionalities and sizes. The findings of the experiments demonstrate that the suggested hyper-heuristic, in most cases, produces higher-quality modularizations in less time than the compared algorithms. The average numerical improvement of the proposed algorithm in terms of modularization quality and execution time on ten folders of Mozilla Firefox is 77.607 and 59.448%, respectively.

Object detection plays a crucial role in identifying and pinpointing objects within images or videos, gaining popularity for its diverse applications in recent times. This document examines the latest progress in object recognition through deep learning. It includes an examination of benchmark datasets and evaluation standards commonly employed in recognition tasks, along with an exploration of key architectures addressing object recognition challenges. Additionally, a scrutiny of contemporary lightweight classification models is provided. To conclude, a standard comparison of the state-of-the-art approaches from different aspects is conducted across various criteria.

Humans perceive and understand the surrounding world through their eyes and analysis with brains. Computer vision endeavors to equip computers with the ability to see images and and process it akin to human vision. Our objective is to develop a model capable of accurately identifying handwritten digits from their images. This will be achieved through the utilization of Convolutional Neural Network principles and its implementation  on the MNIST dataset. The primary aim is to grasp and implement Convolutional Neural Network concepts effectively. While the focus is on creating a model for digit recognition, the potential expansion to letters and individual handwriting is considered. The outcomes of our experiments exhibit a notable enhancement compared to prior works, with an accuracy of 99.30%, signifying the success of the proposed method in recognizing handwritten digits.

The advancement of technology and the emergence of multi-objective optimization problems in various scientific domains have led to the research and presentation of new meta-heuristic algorithms to solve such problems. Although these algorithms have been able to find a relatively good approximation of the optimal Pareto front, but a complete optimization has not been carried out yet. In this paper, to increase the optimality of the generated Pareto front, we present a multi-objective version of the city council evolution algorithm (CCE) called the multi-objective city council evolution algorithm (MOCCE). In the presented algorithm, an archive with a fixed size is considered for storing and retrieving optimal Pareto solutions. This archive is used to define the hierarchical structure of city councils and to simulate its evolution in multi-objective search spaces. The efficiency of MOCCE algorithm has been evaluated on 18 well-known multi-objective test functions known as UF and IMOP and with the results of multi-objective ant lion optimization (MOALO), multi-objective orthogonal mould algorithm (MOSMA) and multi-objective artificial hummingbird optimization algorithms (MOAHA) have been compared. According to the results of the Friedman's mean rank test, in all UF test functions, MOCCE ranks first among all compared algorithms in terms of generation distance (GD), inverse generation distance (IGD) and maximum spread (MS) criteria. Also, this algorithm takes the first rank in all IMOP test functions in terms of GD criterion and the second rank in terms of IGD and MS criteria

In this article, we present a strategy with the aim of increasing the accuracy of early detection of buyers from customers browsing in an online store. The main advantage of this detection is reducing the risk of losing users with high purchase probability and increasing the conversion rate. Due to the consistency of the prediction and diagnosis framework in data mining techniques, the focus of this article is on the optimal use of pre-processing methods, with the aim of improving the quality of input data to classification algorithms. For this reason, in the proposed strategy, we use a set of algorithms for converting nominal content into numerical, normalization, outlier data detection, feature selection and balancing. Then, we give the modified data to a set of different classification algorithms, including C4.5 decision tree and multi-layer perceptron, and combined classification algorithms of random forest, bagging and gradient boosting. The evaluation of the results shows that the highest amount of accuracy obtained in this research by using ensemble classifiers has reached 94.42%, which compared with the best results of previous works, the accuracy of diagnosis has increased.

Todays, artificial intelligence methods can extract the knowledge hidden in the educational data sets by discovering the relationship between different features. This knowledge can help educational systems in making better decisions and having more advanced plans to improve the academic performance of students. The aim of this study was to identify the factors affecting the academic progress of students and to use a technique that can predict the academic progress of students with the highest percentage of accuracy. Accordingly, artificial intelligence methods including Support vector machine, K-nearest neighbor, Decision trees and Random forest have been applied. Finally, the models was evaluated using Accuracy, Precision, F-measure, Sensitivity, Specificity and Classification error. The results showed that the support vector machine had the best performance in predicting the academic progress of students.

The purpose of this research is to provide a structured method to evaluate the performance of researchers in research organizations using data coverage analysis based on objective, accurate, and precise indicators. To do this, after preparing a questionnaire by studying the literature and interviewing elites, the data collected from the researchers of a research center has been analyzed, and the input and output indicators affecting the efficiency of the researchers have been determined with the help of the data envelopment analysis method. The proposed method in a research center on 56 researchers with four input indicators and ten output indicators, implementation, and results show that the efficiency score of 11 researchers was one, and other researchers had scores less than one, and their efficiency was measured compared to the first 11 people.

With the acceleration of the development of science and the publication of articles and the increase of scientific fields, finding suitable research partners, finding research sources and research fields for researchers and relevant institutions is becoming more and more difficult. By choosing these things correctly, you can get the most efficiency from the cost and time spent on research. To solve this problem, a scientific network can be created by creating a network including articles, scientists, and other scientific entities and the connections between them, and predicting the connections that will be formed in the future using link prediction. In this paper, a framework based on machine learning is presented for link prediction in scientific networks. In this framework, by weighting the network based on time and content, calculating embedded structural and textual features, feature selection and extraction, and finally negative sampling using clustering, a machine learning model is trained for link prediction. Each of the steps of this framework was tested separately and all together, and the results showed that the proposed weighting method for the network of references and authors' collaboration increases the accuracy of the weighted similarity criteria and, as a result, increases the accuracy of the entire algorithm. Also, negative sampling using clustering makes the machine learning algorithm better trained. The textual features of scientific data such as the title and abstract of articles also play an effective role in predicting future links.

The least square error method yields acceptable results in system identification. The process of parameter estimation helps solve the linear equation . The important point is that the normal method of solving the above problem has a computational complexity for an matrix. In solving this problem, the computational complexity increases with the increase of (the size of the data matrix). On the other hand, availability of more samples leads to better modeling of the system. In the practical problems of system identification, when the number of input data is large the computational complexity increases greatly. Recently, in the field of quantum computing, some algorithms have been presented that significantly reduce the computational complexity. Among the most important of them is the HHL algorithm, which solves the above linear equation in time, where is the condition number and s is the sparsity of the matrix. In this article, the goal is to present the developed quantum algorithm for solving the problem of least square error identification (GLS method). In this article, two classical-quantum and all-quantum methods are presented. Unlike conventional HHL methods, the proposed methods in this article are able to calculate unbiased parameters with non-Hermitian matrices, and color noise. The proposed classical-quantum method has a computational complexity of of and the all-quantum method has an order of in relation to the size of the data matrix. The results and comparisons show that the methods proposed in the article have less complexity and limitations than classical methods such as psudeo-inverse complexity).

Software testing is an important and necessary matter in the development process of any software. Nowadays, an important part of software testing carrying out automatically. In automatic software testing, there may be obstacles that make the test difficult. One of these obstacles is the infeasible paths in the control flow graph obtained from the codes of software under test. The existence of such paths can cause problems in various fields such as test data generation, security and etc. and also cause resources waste. In order to prevent the mentioned problems, detection of infeasible paths can be very critical. But in some cases, recognizing these paths can be a time-consuming task due to the large amount of codes. In this article, a method is proposed that can increase the infeasible paths detection speed by applying reinforcement learning algorithms. In this method, the speed of the Z3 SMT solver has been increased by using the DQN agent. As the speed of the Z3 tool increases, the speed of all algorithms and methods that use this tool to detect infeasible paths in control flow graphs will also increase. Also, the Z3 tool is using in various other fields, and improving its performance in those fields can also be effective. The proposed method in this article has been tested on QF_NRA and QF_NIA logics. In the experiments, it is shown that using the proposed method, the speed of the Z3 solver can be increased up to 4.7 times.

Demand for quality products has been increasing for decades and is now increasing. Quality control ensures increased product production using an automated, cost-effective and non-destructive method. In the last few years, significant results have been achieved in various agricultural and food sectors. Despite having different programs and many different machine learning techniques, this study only explains the statistical technologies of machine learning with machine vision systems in agriculture due to the wide range of machine learning programs. Two types of machine learning techniques, such as supervised and unsupervised learning, have been used for agriculture. In this research, software solutions rely on image processing techniques such as: artificial neural networks, genetic algorithm, deep learning and fuzzy logic for automatic detection as well as classification of different degrees of fruit. According to the investigations, it was shown that fuzzy logic with 80 to 93%, 96.22% and 91.2% accuracy respectively for grading raisins based on qualitative characteristics, calculating the mass of almond kernels without breaking, grade Apple classification, also neural network has been used with 90%, 99% and 90.4% accuracy respectively to detect apple leaf disease, modelling drying kinetics of strawberry slices and detecting the severity of tomato disease using deep learning. There is also more reference to the study and description of product classification methods that using the mentioned algorithms and their relationship with the software can be a big step in quality classification of products.

In the numerical analysis of fluid mechanics problems, especially in high resolution simulation, the reduction of computational costs has always been of great importance. The use of reduced order models, which will increase the speed of computation by reducing the constraints of the original model. It is a suitable surrogate model for the original governing equation. In this research, using dynamic mode decomposition and based on principles of dynamical systems, the governing equation has been converted to a reduced order model. The results show if the Reynolds number increases and the effects of the viscous term in the governing equation are reduced, the required dissipation in the surrogate model to stabilize the numerical solution is decreased. Also, due to the incompleteness of the modal space and removing the effects of some modes, the instability will be enhanced.