In this paper we show that corpus-level aggregation hinders considerably the capability of lexical metrics to accurately evaluate machine translation (MT) systems. With empirical experiments we demonstrate that averaging individual segment-level scores can make metrics such as BLEU and chrF correlate much stronger with human judgements and make them behave considerably more similar to neural metrics such as COMET and BLEURT. We show that this difference exists because corpus- and segment-level aggregation differs considerably owing to the classical average of ratio versus ratio of averages Mathematical problem. Moreover, as we also show, such difference affects considerably the statistical robustness of corpus-level aggregation. Considering that neural metrics currently only cover a small set of sufficiently-resourced languages, the results in this paper can help make the evaluation of MT systems for low-resource languages more trustworthy.

In this paper we study the fine-tuning of pre-trained large high-resource language models (LLMs) into many-to-one multilingual machine translators for extremely-low-resource languages such as endangered Indigenous languages. We explore those issues using datasets created from pseudo-parallel translations to English of The Bible written in 39 Brazilian Indigenous languages using mBART50 and WMT19 as pre-trained models and multiple translation metrics. We examine bilingual and multilingual models and show that, according to machine translation metrics, same-linguistic family models tend to perform best. However, we also found that many-to-one multilingual systems have a tendency to learn a "rogue" strategy of storing output strings from the training data in the LLM structure and retrieving them instead of performing actual translations. We show that rephrasing the output of the training samples seems to solve the problem.

In this paper we present PeLLE, a family of large language models based on the RoBERTa architecture, for Brazilian Portuguese, trained on curated, open data from the Carolina corpus. Aiming at reproducible results, we describe details of the pretraining of the models. We also evaluate PeLLE models against a set of existing multilingual and PT-BR refined pretrained Transformer-based LLM encoders, contrasting performance of large versus smaller-but-curated pretrained models in several downstream tasks. We conclude that several tasks perform better with larger models, but some tasks benefit from smaller-but-curated data in its pretraining.

In robotics, precise models are critical for ensuring safety and functionality. However, acquiring a precise model characterizing a system's dynamics can be challenging. One of the alternatives to address this issue is system identification, which aims to obtain models through physical and experimental observations. In this manner, the developments in machine learning algorithms, such as neural networks, have significantly improved the modeling of complex and nonlinear phenomena. In this work, a mass-spring-damper (MSD) system and a low-cost original elastomer-based Series Elastic Actuators (eSEA) assembly are used to evaluate the performance of system identification models. The black-box models selected are variations of the AutoRegressive Moving Average with eXogenous input (ARMAX) algorithm. The gray-box model aims to estimate the parameters of 4 friction models; the optimization is done utilizing physics-informed neural networks (PINNs). For both case studies, the PINNs outperformed the black-box models. In the didactic example, the parameters obtained are close to the ground truth, and the highest determinant coefficient obtained is 0.99. The friction model that best represents the robotic actuator is the LuGre model, with the parameters obtained using the PINNs, outperforming the best black-box model by lowering the mean absolute error (MAE) by 30.83%. With a determinant coefficient of 0.94, the model shows a high capacity for describing the multiple nonlinearities present in the system. Copyright (C) 2024 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)

In the field of nondestructive testing and structural health monitoring, ultrasonic waves are widely utilized to identify defects and characterize materials. Recently, data-driven machine learning models have been proposed for strain estimation using shallow-models and Principal Component Analysis (PCA). However, little research effort has been guided towards the development of real-time strain estimation hardware accelerators. This study presents a novel comparative analysis of hardware implementations of PCA on a low-cost SoC-FPGA using High-level Synthesis (HLS) and HDL-based architectures. The comparison was conducted in terms of relevant metrics: hardware occupation, latency, and computational efficiency. Additionally, we demonstrate a scalability analysis considering floating-point bit-width representation and the number of operators. The proposed HDL-based architecture was able to achieve similar performance in comparison with the HLS-based implementation. The advantages of the proposed hardware accelerators are shown by their real-time inference capabilities, low power consumption, and reduced hardware utilization associated with low latency and elevated computational efficiency.

The present work proposes i) the extraction of the dynamics of a liquid-saturated steam heat exchanger through a black-box approach, considering multiobjective optimization of an Echo State Network (ESN) hyperparameters based on performance and complexity metrics and ii) the application of a fuzzy control on the ESN obtained to follow a temperature trajectory. Results showed that the generated model can be controlled and, when compared with a model obtained in an approach with only performance optimization, a similar performance is achieved but with a significant model complexity reduction.

The characterization of hysteretic components poses a difficult nonlinear system identification problem. Several studies have addressed this by employing artificial neural networks, where deep learning (DL) has recently gained attention in system identification tasks. However, there is a lack of studies comparing different deep neural network (DNN) architectures. Therefore, this work proposes the comparison of three DNN architectures, including feedforward neural networks (FFNN), long short term memory (LSTM), and convolutional neural networks (CNN), for the characterization of a piezoelectric positioning system (positioner) typified by hysteresis. Moreover, Bayesian optimization is employed for hyperparameter tuning in all DNN architectures. Results show that all DL architectures achieved desirable values for the coefficient of determination (R2) and root mean squared error (RMSE). However, LSTM obtains the best overall results, outperforming both the FFNN and CNN, being a more appropriate black-box architecture for identifying frequency-dependent hysteresis loop shapes.

Illegal tapping of fuel pipelines has recently become one of the most relevant safety problems faced by the industry. Hundreds of illegal interventions have been reported around the world, causing a significant number of deaths, relevant impacts on the environment, and capital loss. Therefore, it is important to develop systems that are able to detect such scenarios at an early stage, enabling a fast counteract. To this end, machine learning algorithms can train models on available data for detecting future issues. Most recently, ensemble learning and dynamic classifier selection (DCS) techniques have been achieving promising results in supervised learning tasks. Such models are usually trained based on a single criterion. However, it is desirable to take into account both the number of false positives (FP) and false negatives (FN) for the illegal tapping detection task, since they are conflicting and both lead to financial losses and/or accidents. Therefore, this work proposes a novel DCS technique based on multiple criteria, namely overall local class-specific accuracy (OLCA), which employs multi-criteria decision making for dynamically selecting the best classifier for a new sample given the local true positive and negative ratios. A numerical experiment is conducted for assessing the generalization performance of the proposed method in an oil pipeline, with the goal of detecting illegal taping using pressure transient signals. Results show that OLCA is able to reduce the number of both FP and FN when dynamically selecting the classifiers of a baseline Random Forest ensemble.

Artificial neural networks (ANN) has attracted attention of the academic community by the current progress that this technique has provided in speech recognition and digital media such as as image, video, audio, and signal processing. Some fields, as industrial process control and product development can be highly benefited by the development of techniques based on the proven potentialities of ANN models, allowing more accurate simulation, better adaptation to changing environments, and greater robustness in model-based fault diagnosis. Along with the advance of ANNs, there is a trend of open-source softwares use for soft computing which facilitates the access of the interested readers to implement their own codes and to explore other applications. Historically evolutionary algorithms such as the Genetic Algorithm (GA) have been implemented to evolve the architectures to search for solutions, in order to solve this fundamental issue that is still an open problem in the general case. Therefore, the present paper investigates the application of ANN to model the nonlinear aircraft turbojet engine through black-box approach. For that purpose it was used real-world measurements of aircraft engine's fuel and rotation as input and output, respectively. In order to facilitate the design, the ANN was optimized aiming to determine the best topology according to the one-step-ahead and free-run simulation. The results obtained encourage the use of automatically generated ANN architectures for dynamic system modeling.

The proportional-integral-derivative (PID) controller is widely adopted to control numerous process in industrial applications owing to its simplicity, clear functionality, and effectiveness. Due to a wide range of applications, distinct methods to determine PID gains are available in the specialized literature, and multiobjective optimization methods became an attractive approach to solve this problem, mainly when more complex systems are considered. Additionally, it is important to compare novel metaheuristics with traditional methods to follow the evolution of the state of the art. For that purpose, it was proposed a new multiobjective formulation of a PID-controlled antilock braking system (ABS) optimization through nondominated sorting genetic algorithm II (NSGA-II), multiobjective dragonfly algorithm (MODA), multiobjective salp swarm algorithm (MSSA) and a new version of MSSA with opposition based learning initialization and evolution. The experimental results were compared through spacing, euclidean distance and hypervolume metrics and showed that the change made in the MSSA (1) have improved the dominance and (2) the spread of Pareto front (PF) solutions.