We feel gratified to present this topical issue, a bouquet of extended versions of a carefully selected subset of articles as accepted, registered, presented and appeared in the proceedings of ISMSI20 & ISCMI20, the twin annual conferences of IICCI held in the year 2020.

The calendars have flipped and we are now in 2024. However, the year 2020 continues to loom large. In fact, 2020 will remain a nightmare and an unforgettable experience for all of us who survived the Covid-19 onslaught. The dark clouds of uncertainty which started hovering around Wuhan in January 2020, rapidly spread to the rest of the People’s Republic of China and ultimately engulfed the entire world, which eventually snowballed into an unparalleled turmoil and universal crisis.

Since then, as a pandemic-ravaged world was struggling to come to grips with skepticism all around, the human lives—both personal as well as professional, had undergone a massive upheaval. Nevertheless, it was heartening for us to observe that notwithstanding the global headwinds, the like-minded scientific community remained upbeat. Instead of responding through knee jerk reactions and looking for greener pastures, peers had dealt with Covid-19 catastrophe with admirable and astonishing resilience. They continued to navigate the challenges and harness the hidden potentials that Computational Intelligence possesses. The increasing number of submissions for both—ISMSI and ISCMI—in the year 2020, clearly indicated the sustained appetite for continuing investigations by the researchers. Their active participations and encouragements surely made our resolve and commitments stronger to move ahead and helped in avoiding derailing our journey during those traumatic days.

Professional success is a dream that lives in all of us. For researchers, it is not the monetary gain that always drives them. Instead, the recognitions and their value to the society that they crave for. From that perspectives, getting noticed and maximizing the impact of one’s work on the society by being cited are challenging but essential ingredients for tasting a sense of accomplishments in the scientific world. Toward that, publishing in a specific special issue of a leading journal facilitates in attracting more visibility and probably more citations since the peers always prefer and look for specific topics of their interests. In order to enable our conferences’ authors to present their research more impressively and more elaborately which would strengthen their outreach as well as impact, we, by following in the footsteps of the previous years, selected a subset of the papers (42) of both the conferences of 2020, based on the novelty and having scope of enough extension. The authors pertaining to those 42 articles were communicated and advised to expand (at least, by 50%) their respective conference manuscripts and submit for possible publications at this topical issue of NCAA.

All the submissions were reviewed by no fewer than two independent domain experts as per the prevailing protocol of NCAA. The emphasis was on the quality, originality and creativity of each manuscript. Finally, 17 articles of the authors encompassing Asia, Africa, Australia, Europe and North America secured the green signal of the Guest Editors as well as the Editor-in-Chief for ultimate publications in this topical issue. We now present the outlines of each one of the 17 published papers.

Debnath el al. formulated text summarization as a multi-objective optimization problem using modified cat swarm optimization (MCSO). The authors represent each population as feasible individuals with summary length constraints, evaluated by objectives such as "coverage and informativeness" and "anti-redundancy." On the DUC-2001 and DUC-2002 datasets, the approach showed significant improvements in ROUGE scores and produced readable, concise and relevant summaries efficiently.

Sakai and Tamiya introduced a new fixed point format, shifted dynamic fixed point (S-DFP), which adjusts the data representation range by adding bias to the exponent, resulting in improved accuracy during quantized neural network training, as demonstrated on the ImageNet task with ResNet models; for instance, the accuracy of quantized ResNet-152 was increased from 76.6% with conventional 8-bit DFP to 77.6% with 8-bit S-DFP.

Kerim and Genc investigated the relationships between specific attributes of mobile games and their success. Over 17,000 games were analyzed to determine factors that contribute to success. The study found that In-App Purchases (IAPs), genre, number of supported languages, developer profile and release month significantly impact the success of a mobile game. A novel success score was developed to measure multiple objectives.

Bhowmick et al. proposed the Character Inclusion Transformer, an improved transformer network that learns character-level information from code-mixed sentences, enhancing performance with low-resource datasets. Additionally, two transfer learning architectures using the mBERT pre-trained model are introduced. The study evaluates three NLP tasks on six datasets in terms of weighted and macro-average precision, recall and F1 scores.

Davendra et al. introduced a novel chaotic flower pollination algorithm (CFPA) for addressing a tardiness-constrained flow-shop scheduling problem with simultaneously loaded stations. Traditional solutions have relied on standard deterministic algorithms, but they have developed a metaheuristic approach using the flower pollination algorithm enhanced with chaos maps for added stochasticity. The proposed approach has been demonstrated with fifteen experiments and thirty scenarios simulating industrial conditions.

Dasari et al. addressed the class-imbalance issue in additive manufacturing (AM) process data, where defect-free data vastly outnumber defect data, leading to sub-optimal classification results. To improve defect classification essential for quality inspection, the study proposes a cluster-based adaptive data augmentation (CADA) method for oversampling the minority class. Quantitative experiments conducted using AM datasets from the aerospace industry and a publicly available casting dataset demonstrate that CADA outperforms random oversampling and the SMOTE method.

Khumalo et al. explored the potential of quantum technology to address the intractability of solving NP-Hard Combinatorial Optimization Problems (COP), specifically the Travelling Salesman Problem and the Quadratic Assignment Problem. It compares two classical optimization methods, Branch and Bound and Simulated Annealing, with two quantum optimization methods, Variational Quantum Eigensolver (VQE) and Quantum Approximate Optimization Algorithm (QAOA), executed on classical devices and IBM’s Noisy Intermediate-Scale Quantum (NISQ) devices. Experimental results show that current classical devices outperform NISQ devices, with VQE performing better than QAOA due to fewer required operations.

Perumal and van Zyl addressed the challenge of parameter calibration in agent-based modelling and simulation (ABMS), where increasing parameters lead to computational infeasibility due to the "curse of dimensionality." The authors proposed a comprehensive and adaptive ABMS framework that allows for flexible integration of parameterization strategies and surrogate models to calibrate an infectious disease ABM. The framework’s performance in terms of accuracy and efficiency is evaluated across different strategy–surrogate combinations. Results demonstrate that surrogate-assisted sampling strategies outperform baselines in accuracy. Specifically, the Metric Stochastic Response Surface strategy with Support Vector Machine (SVM) surrogate is identified as the best for obtaining actual synthetic parameters.

Saleem and Kovari explored the impact of sampling rate on the accuracy of online signature verification systems. Unlike previous approaches based on interpolation techniques, this research investigates the effects of different sampling rates without interpolation. The study proposes a novel online signature verification system that adapts the sampling frequency based on the signer. Twenty verifier configurations were tested on various public signature databases using different preprocessing methods and sampling rates. The results indicate that there is an optimal range of sampling frequency (15–50 Hz) and signature point count (60–240) that minimizes error rates.

Helbig and Engelbrecht extended a previous study on addressing many-objective optimization problems (MaOPs) by incorporating the partial dominance relation into another multi-objective optimization algorithm (MOA). The performance of the partial dominance relation is evaluated against both the original MOAs and state-of-the-art algorithms, demonstrating its efficiency in solving MaOPs. The findings provide further evidence that the partial dominance relation is a valuable approach for addressing the challenges of many-objective optimization problems.

Tsuji et al. proposed a greedy search algorithm for neural network partial quantization that does not require re-training. The algorithm achieves practical combinations of quantization layers with low computational complexity. Experimental results demonstrate that the proposed method achieves significant model size compression (e.g., 4.2 × in ResNet50) with minimal accuracy degradation (e.g., 0.03% in ResNet50).

Mathonsi and van Zyl proposed an approach to overcome multivariate anomaly detection challenges and benchmarked against well-established models. Three deep learning architectures, namely cascaded neural networks, reservoir computing and long short-term memory recurrent neural networks, are investigated. The results indicate that deep learning outperforms or is at least competitive with the classic models in anomaly detection tasks.

Olukanmi et al. addressed the drawback of local minima in the k-means clustering algorithm by proposing a technique to compare different initializations directly. By comparing the initializations based on the minimum inter-center distance (MIND), the need for repeated runs of the algorithm can be eliminated. The proposed technique also serves as a general approach for optimizing k-means seeding algorithms. Two popular algorithms, k-means and k-means++, are optimized using this technique.

Stander et al. contributed to the field of chemical plant systems optimization by exploring the use of machine learning models as surrogates for computationally expensive simulation models. The study extends previous research on surrogate-assisted genetic algorithms (SA-GA) and introduces a novel multivariate extension called surrogate-assisted NSGA (SA-NSGA) for more complex plant design and operation optimization problems. The proposed approach is evaluated on the pressure swing adsorption (PSA) system, and extensive experimentation compares different meta-heuristic optimization techniques and machine learning models as surrogates.

Hintze and Adami examined the influence of training methods on information propagation in artificial neural networks (ANNs). ANNs are inspired by natural brains but differ in connectivity and training methods. The study measures transfer entropy, which represents information transfer between groups of neurons, to compare ANNs trained with backpropagation and neuroevolution. The results show that while the distribution of connection weights is similar, neuroevolution leads to networks with more focused information transfer among small groups of neurons and greater robustness to weight perturbations compared to backpropagation.

Huo and van Zyl addressed the problem of catastrophic forgetting in neural networks during incremental learning, specifically focusing on similarity-based loss functions. Four well-known loss functions (Angular, Contrastive, Center and Triplet) were examined, and their impact on catastrophic forgetting was evaluated across multiple datasets. The results indicate that the rate of catastrophic forgetting varies among the loss functions, with Angular loss being the least affected, followed by Contrastive, Triplet loss and Center loss with good mining techniques.

Lin and Sheu proposed particle swarm optimization (PSO) and simulated annealing (SA) algorithms to find the maximum retardance in dextran–citrate-coated ferrofluids. The optimized parameter combinations were determined, and the corresponding maximum retardances were achieved. The PSO algorithm was found to be more effective than SA in optimizing retardance. Furthermore, a hybrid optimization approach combining PSO with sequential quadratic programming (SQP) algorithm was proposed for global solution exploration. The study provides insights into two-level hybrid optimizations for the global exploration of retardance in dextran and citrate-coated ferrofluids.

Hope this topical issue will turn out to be a catalyst for the advancement of the state of the art of Computational Intelligence. The manuscripts, as appearing in this issue, dwell on many challenging topics of vital importance, and these should hold immense significance in shaping the future of CI.

As we are now all set to unveil this topical issue, it seems appropriate for us to assert that the feedback of the readers is very crucial, which we are perpetually soliciting in order to leverage our strengths and address our lacunae for processing the subsequent topical issues in this series. If this issue as a whole is able to gain traction with the peers and the researchers find the collection of articles in the present issue informative and handy for furthering their investigations, we will consider our efforts paid off in the end.