نشریه محاسبات نرم
سال یازدهم شماره 2 (پیاپی 22، پاییز و زمستان 1401)

While many studies estimate air pollutants such as particulate matter (PM), PM2.5 and PM10, have used aerosol optical depth (AOD) products from satellite sensors, utilizing these products for mapping pollution in smaller cities like Tehran is not effective due to their coarse resolution. To address this problem, this study directly uses the Level 1 products of MODIS (instead of aerosol and AOD products). The proposed method employs a deep neural network and a random forest model to estimate the PM values using data from the first two bands of MODIS. The results show the superior performance of the proposed models compared to some state-of-the-art PM estimation methods in recent years. The outcome of this research is the development of a PM map generation software for Tehran (mapping PM2.5 and PM10 concentrations) using freely available MODIS images.

Providing an accurate estimate of the required cost of software development is necessary for the effective management and success of software projects. In the early stages of software development, detailed information about the project is unavailable, making it challenging to provide an accurate cost estimate. With the aim of increasing the accuracy of software development cost estimation, this research has presented a method based on case-based reasoning and Tabu search, which by removing unnecessary features, provides an estimate with the minimum required features for software projects. The initial solution selection mechanism, objective function evaluation index, and missing data removal are the variable parameters of the current research, whose effectiveness has been investigated in the form of 8 different scenarios. The predictive power of the presented method has been measured by the comprehensive and non-comprehensive cross-validation method in the Albrecht, China, Desharnais, Kemmerer, Kitchenham, Maxwell and Miyazaki datasets and evaluated using the two methods of Wu and Liu. The results have shown the superiority of the presented method over Wu and Liu's methods based on the evaluation indicators of the prediction level, median relative error, mean relative error, and mean absolute error. In addition, the effectiveness of the presented method is independent of the methodology and environment used for software development, which by focusing on the general knowledge part of case-based reasoning, has caused the stability of the accuracy of the estimates presented in the research data set.

Authentication in cyberspace aims to maintain security, and if it is done with high accuracy, it can be considered for the continuity of service delivery under different conditions. The present research aims to propose a fully applicable method for authentication of the applicants for electronic services in real time. In order to prevent the possible tricks of the users, in the proposed method, the identification of facial muscle movements and iris biometric measurement have been used. The iris creates more reliability and cannot be stolen or faked; because it must be available live. A method based on two-stream 3D deep learning is proposed for authentication and simultaneously identifying facial muscle movements and distinguishing a living person from an image. According to the evaluations, it was found that the proposed method provides significant assurance for public use and is applicable in real and practical conditions. Using the proposed method, the accuracy is 99.99%, and the average precision of authentication and identifying people in both CASME and CASME2 databases is more than 99.5%.

A person's personality is a set of characteristics and reactions that cause a person to behave in different situations. Usually, these behaviors are repeated in similar situations. Having information about a person's personality allows us to predict a person's reaction in similar situations. On the other hand, considering that people in virtual social networks do not feel direct control over their opinions, using these networks to collect information and analyze people's personalities is very significant and logical. In this article, by receiving information from users 'social networks and comparing it with a questionnaire filled in by individuals, we test the correctness of answering the questionnaire questions and analyze users' personalities in a more accurate way.  By analyzing users' opinions through intelligent text mining techniques and machine learning, we aim to predict the personality of users. Finally, in all four methods of the SVM algorithm as well as the deep neural network method, the criteria of accuracy, accuracy, and readability were relatively high (between 0.75 and 0.99), indicating the high correspondence of the opinions collected through the questionnaire with the posts. Published in cyberspace are the same people.

Nowadays, meta-heuristic optimization algorithms have become very popular in solving optimization problems. By using this group of algorithms, many engineering problems can be solved easily and away from complexity. The Nephron-2 Optimization Algorithm (NOA-2) is one of these algorithms that is the extension of the first version of Nephron Algorithm Optimization. This algorithm is inspired by the functioning of the nephron in the human kidney. The structure of the NOA-2 algorithm proposed in this article according to the behavior of the nephron consists of 4 parts: Separation, absorption, transpiration, and excretion. In order to evaluate the performance, the results of the NAO-2 algorithm and five other famous optimization algorithms on seven optimization problems have been investigated. In this evaluation, two measures of solution quality (objective function) and computational solution time are considered for evaluation and comparison. The results show that the NAO-2 algorithm found the best objective function in a certain time compared to other algorithms and also obtained the optimal solution of the seven studied problems in less time than other algorithms

Nowadays, the number of applications that require a short response time is increasing, and utilizing fog environment has recently received a lot of attention. Due to the dynamics of the resource usage pattern in most Internet of Things applications, a fixed location cannot be considered for the placement and execution of services in the fog environment, and therefore the services must be dynamically placed in the fog environment. The problem of placing required Internet of Things services in cloud devices with limited resources is known as an NP-hard problem. In this article, a dynamic method based on multi-objective genetic algorithm with non-dominated ranking is presented to solve this problem. In the proposed method, learning automata are used to improve genetic behavior and dynamically adjust mutation and crossover rates. The proposed method is simulated using iFogsim and the simulation results show that the proposed method has a better efficiency than the compared algorithms by simultaneously considering the three criteria of service delay, cost and energy consumption. In terms of cost, compared to the two CSA and LRFC methods, it has decreased by 11% and 21%, respectively. Also, in terms of the average service delay, compared to the CSA and HAFA methods, the proposed method has decreased by 7% and 15 %, respectively. In terms of energy consumption, the proposed method shows an improvement of at least 8% compared to other methods.

The prevalence of infectious diseases in the community depends on various factors, including the severity of the disease, compliance with health and communication protocols, vaccination rate, impact factor and duration of its effectiveness on immunization, and effectiveness of treatment protocols and average duration of treatment. Having a model based on which the behavior of the disease can be predicted according to various parameters, can help community leaders to deal with these types of diseases. In the studies conducted so far in this field, the impact of some factors such as compliance with health guidelines, the latent period of the disease and immunity after the disease at the same time as the main factors such as the prevalence rate and vaccination rate have not been fully considered. In this paper, a new and comprehensive model based on Markov theory is presented for predicting the behavior of the Covid-19 disease. This model can imitate the behavior of the disease in different conditions by receiving the given parameters. Numerous simulations with different values of input parameters and their similarity to the actual behavior of Covid-19 disease show the accuracy of the model.

In recent years, deep learning has become the main motive of innovative solutions to artificial intelligence problems, which is made possible by increasing the amount of data available, increasing computing resources, and improving techniques in deep network training. The development and increase of computer processing power and the empowerment of artificial intelligence techniques such as machine learning and deep learning have made it easier for many aerospace projects to be implemented. Theoretical and biological arguments show that in order to build an intelligent system with the ability to extract high-level and powerful representations from data, models with deep architecture that include many nonlinear processing layers are needed. Arguably, the best and most widely used examples of these networks are multilayer neural networks due to their compatibility with data types. Deep neural networks have different structures, different types, and species, and they are used according to the type of data and the purpose of the problem, and each has its strengths and weaknesses. In this paper, the study and application of these networks in various aerospace issues are discussed.

Nowadays, due to people being more willing to shop online through online stores and social media, we are facing the growth of unstructured data like texts on the internet. Hence, text processing and the development of optimal algorithms for extracting knowledge have drawn scholar’s attention to this field. One of the aspects of the text processing field is classifying texts in the form of classes of various sentiments using different algorithms. In this paper, we propose a novel framework to classify the comments based on the user’s sentiment performed in the character-level scenario. Hence, the proposed framework is mounted on the architecture of embedding from the language model triggered by the quad-layer, namely embedding, one-dimensional convolution, the map, and the recurrent neural network. In the proposed framework, first, by using the embedding layer at the level of the character, a constant vector is assigned to them. Next, the semantic and logical relation between the characters per word for surviving word-specific 128-dimensional vectors is extracted by exerting the parallel-oriented one-dimensional convolution operators. After obtaining vectors, based on two recurrent neural network architectures, the relationship between the discovered words and the comment-specific sentiment is determined. The obtained results show that the proposed framework has an Accuracy of 79.87% and a F-score of 79.90% for comments class labeling.

Markov decision processes (MDPs) are used in artificial intelligence and formal verification to model computer systems with nondeterministic and stochastic behaviors. Optimal reachability probabilities and expected accumulated rewards are two main classes of properties that are used in probabilistic model checking. Value iteration and policy iteration are two well-known iterative numerical methods to approximate the optimal values. A main challenge of these approaches is their high running time. In this paper, a new method is proposed for accelerating the convergence to the optimal policy that reduces the running time of the policy iteration method. This method is based on using machine learning to estimate near-optimal policies. For each class of MDP models, we consider several small models for the training and construction step of the classifier model. The classifier model is used to predict the optimal actions of each MDP model. The model is also used to suggest a near-optimal policy for large models of the same MDP class in order to reduce the overall running time. Implementing the proposed method in the PRISM model checker shows a 50% improvement in the average runtime.