In this research, we address bearing fault classification by evaluating three neural network models: 1D Convolutional Neural Network (1D-CNN), CNN-Visual Geometry Group (CNN-VGG), and Long Short-Term Memory (LSTM). Utilizing vibration data, our approach incorporates data augmentation to address the limited availability of fault class data. A significant aspect of our methodology is the application of neural architecture search (NAS), which automates the evolution of network architectures, including hyperparameter tuning, significantly enhancing model training. Our use of early stopping strategies effectively prevents overfitting, ensuring robust model generalization. The results highlight the potential of integrating advanced machine learning models with NAS in bearing fault classification and suggest possibilities for further improvements, particularly in model differentiation for specific fault classes.

With the growing popularity of machine learning (ML), regression problems in many domains are becoming increasingly high-dimensional. Identifying relevant features from a high-dimensional dataset still remains a significant challenge for building highly accurate machine learning models.

Bloat, a well-known phenomenon in Evolutionary Computation, often slows down evolution and complicates the task of interpreting the results. We propose Lexi2, a new selection and bloat-control method, which extends the popular lexicase selection method, by including a tie-breaking step which considers attributes related to the size of the individuals. This new step applies lexicographic parsimony pressure during the selection process and is able to reduce the number of random choices performed by lexicase selection (which happen when more than a single individual correctly solve the selected training cases).

Deep learning is a well suited approach to successfully address image processing and there are several Neural Networks architectures proposed on this research field, one interesting example is the U-net architecture and and its variants. This work proposes to automatically find the best architecture combination from a set of the current most relevant U-net architectures by using a genetic algorithm (GA) applied to solve the Retinal Blood Vessel Segmentation (RVS), which it is relevant to diagnose and cure blindness in diabetes patients. Interestingly, the experimental results show that avoiding human-bias in the design, GA finds novel combinations of U-net architectures, which at first sight seems to be complex but it turns out to be smaller, reaching competitive performance than the manually designed architectures and reducing considerably the computational effort to evolve them.

The advent of the Covid-19 pandemic has resulted in a global crisis making the health systems vulnerable, challenging the research community to find novel approaches to facilitate early detection of infections. This open-up a window of opportunity to exploit machine learning and artificial intelligence techniques to address some of the issues related to this disease. In this work, we address the classification of ten SARS-CoV-2 protein sequences related to Covid-19 using k-mer frequency as features and considering two objectives; classification performance and feature selection. The first set of experiments considered the objectives one at the time, four techniques were used for the feature selection and twelve well known machine learning methods, where three are neural network based for the classification. The second set of experiments considered a multi-objective approach where we tested a well known multi-objective approach Non-dominated Sorting Genetic Algorithm II (NSGA-II), and the Multi-dimensional Archive of Phenotypic Elites (MAP-Elites), which considers quality+diversity containers to guide the search through elite solutions. The experimental results shows that ResNet and PCA is the best combination using single objectives. Whereas, for the mulit-classification, NSGA-II outperforms ME with two out of three classifiers, while ME gets competitive results bringing more diverse set of solutions.

We present Pyramid, a Hierarchical Genetic Algorithm that decomposes problems by first tackling simpler versions of them, before automatically scaling up to more difficult versions while also reducing the population size. Pyramid takes its name from the architectural phenomenon of the Mayan pyramid in Chichen-Itza, which, although constructed from the bottom up, operates in a top down manner through interactions with the sun. This gives a two stage approach: initially we create our pyramid of experiments, with the most complex fitness functions at the bottom, and with increasingly more simplified/decomposed version as we move up through the pyramid. Runs start at the top of the pyramid and populations descend through it, decreasing in size, being exposed to increasingly complex fitness functions, until, at the bottom layer, we have small populations with the original full fitness function. We conduct experiments comparing the performance of Pyramid for a suite of difficult unimodal and multimodal functions. The experimental results show that in three of four cases, Pyramid achieves the same or better fitness scores as standard algorithms, with substantially fewer evaluations, and that in two cases it actually performs statistically significantly better.

Blood vessel extraction in digital retinal images is an important step in medical image analysis for abnormality detection and also obtaining good retinopathy diabetic diagnosis; this is often referred to as the Retinal Blood Vessel Segmentation task and current state-of-the-art approaches all use some form of neural networks. Designing neural network architecture and selecting appropriate hyper-parameters for a specific task is challenging. In recent works, increasingly more complex models are starting to appear, but in this work, we present a simple and small model with a very low number of parameters with good performance compared with the state of the art algorithms. In particular, we choose a standard Genetic Algorithm (GA) for selecting the parameters of the model and we use an expert-designed U-net based model, which has become a very popular tool in image segmentation problems. Experimental results show that GA is able to find a much shorter architecture and acceptable accuracy compared to the U-net manually designed. This finding puts on the right track to be able in the future to implement these models in portable applications

Genetic programming (GP) is an evolutionary computation paradigm for automatic program induction. GP has produced impressive results but it still needs to overcome some practical limitations, particularly its high computational cost, overfitting and excessive code growth. Recently, many researchers have proposed fitness-case sampling methods to overcome some of these problems, with mixed results in several limited tests. This paper presents an extensive comparative study of four fitness-case sampling methods, namely: Interleaved Sampling, Random Interleaved Sampling, Lexicase Selection and Keep-Worst Interleaved Sampling. The algorithms are compared on 11 symbolic regression problems and 11 supervised classification problems, using 10 synthetic benchmarks and 12 real-world data-sets. They are evaluated based on test performance, overfitting and average program size, comparing them with a standard GP search. Comparisons are carried out using non-parametric multigroup tests and post hoc pairwise statistical tests. The experimental results suggest that fitness-case sampling methods are particularly useful for difficult real-world symbolic regression problems, improving performance, reducing overfitting and limiting code growth. On the other hand, it seems that fitness-case sampling cannot improve upon GP performance when considering supervised binary classification.

The objective function is the core element in most search algorithms that are used to solve engineering and scientific problems, referred to as the fitness function in evolutionary computation. Some researchers have attempted to bridge this difference by reducing the need for an explicit fitness function. A noteworthy example is the novelty search (NS) algorithm, that substitutes fitness with a measure of uniqueness, or novelty, that each individual introduces into the search. NS employs the concept of behavioral space, where each individual is described by a domain-specific descriptor that captures the main features of an individual's performance. However, defining a behavioral descriptor is not trivial, and most works with NS have focused on robotics. This paper is an extension of recent attempts to expand the application domain of NS. In particular, it represents the first attempt to apply NS on symbolic regression with genetic programming (GP). The relationship between the proposed NS algorithm and recent semantics-based GP algorithms is explored. Results are encouraging and consistent with recent findings, where NS achieves below average performance on easy problems, and achieves very good performance on hard problems. In summary, this paper presents the first attempt to apply NS on symbolic regression, a continuation of recent research devoted at extending the domain of competence for behavior-based search.

Novelty search (NS) is an open-ended evolutionary algorithm that eliminates the need for an explicit objective function. Instead, NS focuses selective pressure on the search for novel solutions. NS has produced intriguing results in specialized domains, but has not been applied in most machine learning areas. The key component of NS is that each individual is described by the behavior it exhibits, and this description is used to determine how novel each individual is with respect to what the search has produced thus far. However, describing individuals in behavioral space is not trivial, and care must be taken to properly define a descriptor for a particular domain. This paper applies NS to a mainstream pattern analysis area: data clustering. To do so, a descriptor of clustering performance is proposed and tested on several problems, and compared with two control methods, Fuzzy C-means and K-means. Results show that NS can effectively be applied to data clustering in some circumstances. NS performance is quite poor on simple or easy problems, achieving basically random performance. Conversely, as the problems get harder NS performs better, and outperforming the control methods. It seems that the search space exploration induced by NS is fully exploited only when generating good solutions is more challenging.