Preface to the Special Issue on “Advances in Artificial Intelligence: Models, Optimization, and Machine Learning”

Recent advancements in artificial intelligence and machine learning have led to the development of powerful tools for use in problem solving in a wide array of scientific and technical fields. In particular, supervised models allow for the searching, optimization, and classification of data with high complexity, high dimensionality, and vast solution spaces. Problems that had proven challenging or nearly impossible are now solvable given sufficient training time and computational resources, allowing for learning, knowledge discovery, and decision making that easily outperform human abilities. In particular, machine learning models obtained using high-complexity deep neural networks have seen a huge increase in popularity due to their ability to learn functions, rules, and correlations within massive and diverse data sets. Currently, machine learning has demonstrated and is continually demonstrating its potential to solve complex and important real-world problems. Consequently, following a careful and thorough peer-review process, this Special Issue offers valuable contributions to modeling, optimization, classification, and regression for solving problems in a wide variety of technical fields, using modern artificial intelligence and machine learning means.

In the following paragraphs, we provide summaries of these contributions.

Curteanu et al. [1] perform an extensive comparative study between three regression algorithms for the prediction of monomer conversion, numerical average molecular weight, and gravimetrical average molecular weight for the free radical polymerization of methyl methacrylate achieved in a batch bulk process. The first two algorithms are based on the concept of a large margin, typical of support vector machines, but used here for regression in conjunction with an instance-based method, where the learning of problem-specific distance metrics can be achieved either with an evolutionary algorithm or with an approximate differential approach. Another original regression method is based on the idea of denoising autoencoders, i.e., prototype weights and positions are set in such a way as to minimize the error on a slightly corrupted version of the training set.

Leon and Gavrilescu [2] provide a survey of modern methods for prediction and tracking in automotive applications. The paper covers a wide range of methods applied in various scenarios, where pedestrians, vehicles, and other obstacles are found in difficult-to-handle configurations. The scientific contributions analyzed in the survey offer methods using deep neural networks, stochastic methods, motion models, as well as many hybrid approaches to solve problems within scenarios comprising multiple interacting agents with variable or multi-modal behavior, occlusion, high reaction times, and, generally speaking, any contexts encountered within autonomous driving. While the current state of the art generally favors neural-network-based approaches, many non-neural-network solutions are explored as well.

The work by Chou et al. [3] uses an intelligent machine learner to build prediction models with applications in industrial experiments such as resource planning for software projects, the comparison of processor performance, and the estimation of bicycle rentals per day and resources demand for increasing productivity and efficient customer service. The proposed approach matches or obtains better results than the existing methods reported in the literature for the same applications.

The problem of the detection of gender-based violence is tackled by Castorena et al. [4] based on language used on Twitter. The artificial intelligence systems used for this goal are deep neural networks that require minimal preprocessing of data based on feature extraction. The success rate of the proposed method in identifying gender-based violence in the Spanish language in Mexico is about 80%, which is encouraging for future improvements of the proposed method.

Kang [5] proposes an improvement of the k-nearest neighbors (kNN) algorithm using a graph neural network (GNN) to improve the learning process. The resulting contribution is called kNNGNN and consists of generating a GNN that learns kNN rules from a graph representation of the data. The author evaluates both weighted and unweighted versions of kNN using various similarity metrics and demonstrates the applicability of the proposed method for both classification and regression problems.

The work by Muñoz Castañeda et al. [6] describes a new algorithm for the hyper-parameter optimization (HPO) of machine learning algorithms based on the conditional optimization of concave asymptotic functions. It is shown that the size of the data subset does not have a great impact on its performance, and the algorithm only requires an upper bound on the number of iterations to perform.

Drăgoi and Dafinescu [7] review a large number of metaheuristic optimization algorithms inspired by animal behavior, both vertebrates and invertebrates, proposed between 2006 and 2021. The authors note that despite many critiques of the metaheuristic community, the trend of proposing algorithms based on new sources of inspiration remains stable because of the many areas of applicability and the tendency to offer the source code in order to increase the ease of use. Exotic inspiration sources and uncommon behaviors seem to have a greater probability of devising new optimization techniques.

Feng et al. [8] use a model based on the ResNet50 architecture to identify surface defects on rolled strip steel for automotive manufacturing. ResNet50 is combined with other models such as the convolutional block attention module (CBAM) and FcaNet for improved accuracy. The resulting hybrid method is tested using a data set exhibiting defect patterns such as surface scratches, cracks, tears, spots, or oxidation layers. The proposed hybrid model performs slightly better than similar approaches. However, as the authors themselves note, the method requires more computational power than competing lightweight models, considering that the rolled steel coils are evaluated using images acquired in real time at very high rates.

Bădică et al. [9] study hierarchically shaped single-elimination tournaments and propose a dynamic programming algorithm for use in computing optimal tournaments that maximize attractiveness, e.g., where the best players have the chance to meet in the later stages of the competition. The authors also develop more efficient deterministic and sub-optimal stochastic versions of the algorithm.

The goal of the paper by Kušić et al. [10] is to improve artificial intelligence techniques for traffic control by dynamically setting zones with variable speed limits. In addition, method validation is performed using four agents instead of two, as in previous research. This work is important in reducing traffic congestion by automatically adjusting the speed limits and the position of these zones.

In his work, Leon [11] describes the architecture of ActressMAS, a .NET multi-agent framework which allows the implementation of two sub-paradigms in multi-agent systems, i.e., one focused on autonomy and planning, and another focused on interactions and emergent behaviors in agent simulations. Its main advantages are conceptual simplicity and ease of use, which make it particularly suitable for teaching agent-based concepts. However, the framework proves to be sufficiently powerful to implement a large number of algorithms, protocols, and simulations characteristic of intelligent agents and multi-agent systems. The framework and the examples are open-source and publicly available.

The paper by Saeed et al. [12] presents a method to optimize the structure of convolutional neural networks (CNNs) by determining the number of filters and layers for the classification of fingerprints using multiple sensors. This research is important for improving the cost and response time of systems based on CNNs.

Chatterjee et al. [13] propose a method for the automatic identification of plastic bottles from images for recycling purposes. To this end, a model based on a generative adversarial network (GAN) augments a data set consisting of a few original images, while the actual classification is handled by an ensemble based on transfer learning from the InceptionV3 and Xception models. The proposed solution is shown to have very high accuracy, and it is worth mentioning that it seems to handle rotation and translation quite well. However, both training and evaluation are carried out on relatively simple images, each containing a single plastic bottle against a relatively homogeneous background. It would be interesting to see whether in future work the authors will improve their method to handle more diverse and realistic scenarios, such as plastic bottles found among other waste, multiple plastic bottles arranged in piles where they occlude one another, etc.

These 13 papers in this Special Issue have been selected following a process with an acceptance rate of 62%. The authors’ geographical distribution is displayed in

Table 1. Geographic distribution of authors by country.

Country	Number of Authors
China	3
Croatia	4
Germany	2
India	2
Ireland	1
Mexico	6
Romania	12
Saudi Arabia	3
South Korea	3
Spain	3
Taiwan	2
Vietnam	1

, which shows 42 authors from 12 countries.

The guest editors wish to thank the authors for their contributions and for their commitment to improving their work, the reviewers for investing time and effort into analyzing and providing valuable comments and corrections, and last but not least, the editorial staff for managing the review and publication process efficiently and thoroughly. We hope that the selected publications will have a lasting impact on the scientific community and that they will be motivating factors for other researchers to pursue their scientific goals.