Black-box Approach Research Articles

Individualized Models for Glucose Prediction in Type 1 Diabetes: Comparing Black-Box Approaches to a Physiological White-Box One.

Accurate blood glucose (BG) prediction are key in next-generation tools for type 1 diabetes (T1D) management, such as improved decision support systems and advanced closed-loop control. Glucose prediction algorithms commonly rely on black-box models. Large physiological models, successfully adopted for simulation, were little explored for glucose prediction, mostly because their parameters are hard to individualize. In this work, we develop a BG prediction algorithm based on a personalized physiological model inspired by the UVA/Padova T1D Simulator. Then we compare white-box and advanced black-box personalized prediction techniques. A personalized nonlinear physiological model is identified from patient data through a Bayesian approach based on Markov Chain Monte Carlo technique. The individualized model was integrated within a particle filter (PF) to predict future BG concentrations. The black-box methodologies considered are non-parametric models estimated via gaussian regression (NP), three deep learning methods: long-short-term-memory (LSTM), gated recurrent unit (GRU), temporal convolutional networks (TCN), and a recursive autoregressive with exogenous input model (rARX). BG forecasting performances are assessed for several prediction horizons (PH) on 12 individuals with T1D, monitored in free-living conditions under open-loop therapy for 10 weeks. NP models provide the most effective BG predictions by achieving a root mean square error (RMSE), RMSE = 18.99 mg/dL, RMSE = 25.72 mg/dL and RMSE = 31.60 mg/dL, significantly outperforming: LSTM, GRU (for PH = 30 minutes), TCN, rARX, and the proposed physiological model for PH=30, 45 and 60 minutes. Black-box strategies remain preferable for glucose prediction even when compared to a white-box model with sound physiological structure and individualized parameters.

Hierarchical framework for interpretable and specialized deep reinforcement learning-based predictive maintenance

Deep reinforcement learning holds significant potential for application in industrial decision-making, offering a promising alternative to traditional physical models. However, its black-box learning approach presents challenges for real-world and safety-critical systems, as it lacks interpretability and explanations for the derived actions. Moreover, a key research question in deep reinforcement learning is how to focus policy learning on critical decisions within sparse domains. This paper introduces a novel approach that combines probabilistic modeling and reinforcement learning, providing interpretability and addressing these challenges in the context of safety-critical predictive maintenance. The methodology is activated in specific situations identified through the input–output hidden Markov model, such as critical conditions or near-failure scenarios. To mitigate the challenges associated with deep reinforcement learning in safety-critical predictive maintenance, the approach is initialized with a baseline policy using behavioral cloning, requiring minimal interactions with the environment. The effectiveness of this framework is demonstrated through a case study on predictive maintenance for turbofan engines, outperforming previous approaches and baselines, while also providing the added benefit of interpretability. Importantly, while the framework is applied to a specific use case, this paper aims to present a general methodology that can be applied to diverse predictive maintenance applications.

Automatic trajectory adaptation for the control of quality characteristics in a powder compaction process

The manufacturing process of sintered components requires the compaction of powder in a rigid die and the sintering of the resulting green parts. To produce green parts of the same constant quality over a long period of time, regular quality checks and trajectory adjustments are needed to account for changing operating conditions due to e.g. varying temperature, humidity, stroke rate, or powder quality. Also the first setup of the machine requires manual measurements and iterative adjustments of trajectory key points. To overcome these labour-intensive, manual measurements and adjustments, we propose an algorithm for the automatic adaptation of trajectory key-points that for the first time allows bringing and keeping more than one quality characteristic (mass and dimensions) fully automatically within predefined intervals. To safe costs, we further present two real-time capable, model-based solutions by employing grey-box and black-box models for the online estimation of the workpiece mass and height instead of their direct measurement. The grey-box model combines a filling and analytical compaction model and allows estimating mass and height rather than density as typically investigated in literature. The black-box model based on neural networks represents a complete novelty in the field as no similar data-driven approach could be found that aims at estimating mass and dimensions of green compacts. We finally compare the performance of these models to the baseline of sensor measurements as well as study their effect on the overall control performance. Results indicate that for both sensor-based and model-based solutions it is possible to reach and keep aforementioned quality characteristics within desired intervals. While overall the sensor-based approach performed best, also for the model-based approaches a good estimation performance is observed with the black-box model based on neural networks outperforming the grey-box model. While the black-box approach is fully data-driven and thus allows to widely avoid manual modelling activities, it still requires the presence of expensive sensors for capturing training data. The proposed grey-box approach, on the other hand, requires some manual modelling and model calibration activities, but provides a cost-effective solution to the problem and represents an alternative to more expensive, sensor-based implementations.

Interpretable Model-Agnostic Explanations Based on Feature Relationships for High-Performance Computing

In the explainable artificial intelligence (XAI) field, an algorithm or a tool can help people understand how a model makes a decision. And this can help to select important features to reduce computational costs to realize high-performance computing. But existing methods are usually used to visualize important features or highlight active neurons, and few of them show the importance of relationships between features. In recent years, some methods based on a white-box approach have taken relationships between features into account, but most of them can only work on some specific models. Although methods based on a black-box approach can solve the above problems, most of them can only be applied to tabular data or text data instead of image data. To solve these problems, we propose a local interpretable model-agnostic explanation approach based on feature relationships. This approach combines the relationships between features into the interpretation process and then visualizes the interpretation results. Finally, this paper conducts a lot of experiments to evaluate the correctness of relationships between features and evaluates this XAI method in terms of accuracy, fidelity, and consistency.

An explainable two-stage machine learning approach for precipitation forecast

A common post-processing approach to improve precipitation forecasts is to use machine learning models such as artificial neural networks (more specifically, multi-layer perceptrons) as black-box systems. These models utilize different sources of observations or predictors to generate an improved forecast in terms of desired metrics. However, most existing studies employ a single-stage regression model without considering explainability. The small number of studies with two-stage models that combine classification and regression utilize binary classification and still lack explainable artificial intelligence. Therefore, this study proposes a precipitation prediction system which (i) is composed of two stages for better predictions, (ii) compares the utility of binary and multi-class classification over the regression, and (iii) is explainable, unlike prior studies, in that individual predictions of machine learning-based forecasts are interpretable by humans. The proposed two-stage model first estimates the precipitation intensity category using binary or multi-class classification as the first stage and later utilizes precipitation intensity category information in a regression model, which is the second stage, to obtain daily precipitation magnitude. The utilized approach is made humanly interpretable (i.e., explainable) by providing insight into the model-wide importance of predictors and generation processes of the individual predictions (instance-level explanation). The proposed two-stage approach is compared against single-stage and black-box approaches in terms of prediction quality and explainability, where daily station-based observations are used as ground truth datasets. Experiments show that the proposed two-stage approach yields significant improvement (on average, RMSE reduced by 10.50%, and the correlation between numerical precipitation estimates and observed precipitation values increased by 7.5%) compared to the best-performing physical predictor (ECMWF). Analysis of explainability provides insights into the decisions of our two-stage approach, e.g., the usefulness of seasonality-related parameters, multi-class precipitation intensity classification as a first stage, and the predictors for each task (regression or classification).

Black-box Attack against Self-supervised Video Object Segmentation Models with Contrastive Loss

Deep learning models have been proven to be susceptible to malicious adversarial attacks, which manipulate input images to deceive the model into making erroneous decisions. Consequently, the threat posed to these models serves as a poignant reminder of the necessity to focus on the model security of object segmentation algorithms based on deep learning. However, the current landscape of research on adversarial attacks primarily centers around static images, resulting in a dearth of studies on adversarial attacks targeting Video Object Segmentation (VOS) models. Given that a majority of self-supervised VOS models rely on affinity matrices to learn feature representations of video sequences and achieve robust pixel correspondence, our investigation has delved into the impact of adversarial attacks on self-supervised VOS models. In response, we propose an innovative black-box attack method incorporating contrastive loss. This method induces segmentation errors in the model through perturbations in the feature space and the application of a pixel-level loss function. Diverging from conventional gradient-based attack techniques, we adopt an iterative black-box attack strategy that incorporates contrastive loss across the current frame, any two consecutive frames, and multiple frames. Through extensive experimentation conducted on the DAVIS 2016 and DAVIS 2017 datasets using three self-supervised VOS models and one unsupervised VOS model, we unequivocally demonstrate the potent attack efficiency of the black-box approach. Remarkably, theJ&Fmetric value experiences a significant decline of up to 50.08% post-attack.

Design-focused Interpretable Machine Learning Models for Compressive Capacity Prediction of Gusset Plate Connections

Gusset plates can experience large compressive loads when used to connect bracing systems to beam-column joints in steel frames and truss bridges. The strength of a gusset plate is therefore vital to allow the bracing members to reach their ultimate capacity without a brittle failure of the gusset plate connection. Due to the large complexity of the stress distribution mechanism, simplified approaches commonly adopted for design can lead to inaccurate and highly conservative results. The present study aims at developing a simple yet efficient gusset plate compressive strength prediction model that does neither rely on crude assumptions nor require complicated calculations. A meticulously compiled database consisting of 68 experimental and 184 numerical datasets was deployed to aid in developing a high-performance machine learning (ML) model selected amongst 10 machine learning techniques studied. The superiority of the CatBoost and XGBoost models was confirmed with an accuracy of 95% and 96%, respectively significantly outperforming current design practices and widely used empirical formulas. SHapley Additive exPlanations (SHAP) and Taylor diagram were used to explain the feature importance and model performance. The developed interpretable ML models can instill trust in a model that explains its decisions, as opposed to other black box approaches. It was found that CatBoost was the most efficient in predicting the compressive strength of gusset plates with lowest mean absolute error of 50.76 kN and Nash-Sutcliffe model efficiency coefficient of 0.984, compared to the most accurate predictions observed among the existing methods (modified Thornton method) with mean absolute error and Nash-Sutcliffe model efficiency coefficient of 227.22 kN and 0.872, respectively. Finally, a reliability analysis was performed according to the LRFD approach to identify the level of safety of predictions obtained by the CatBoost and XGBoost models. The CatBoost model exhibited remarkably low dispersion with a safety factor of 0.753 thereby providing an adequate level of safety for the design of gusset plates.

Detection of backdoor attacks using targeted universal adversarial perturbations for deep neural networks

Backdoor attacks on deep neural networks (DNNs) using targeted universal adversarial perturbations (TUAPs) do not require training datasets and model tampering, and triggers based on TUAPs can make DNNs output any class the adversary wants. Retraining DNNs using adversarial training for security is time-consuming and does not apply to DNNs in runtime. We want to detect backdoors using a black-box testing approach. We observe that after superimposing random noise on the input of a backdoor attack, the output still tends to remain the same, so we propose Sequential Analysis method based on the Metamorphosis Testing (SAMT). We designed two metamorphic relations for test case generation. Using sequential sampling, we calculate the label stability rate (LSR) and infer whether the image to be verified contains a trigger based on the sequential probability ratio change. The experimental results show that our method has a higher backdoor detection success rate (dsr) than the state-of-the-art detection algorithms. Moreover, our method does not need to use the model structure of DNNs, which has more adaptability and generalization ability. Based on our proposed method, we can simply add a backdoor detection layer to detect backdoors as early as possible, which can eventually alleviate the harm of such backdoors.

A comparison of physics-informed data-driven modeling architectures for ship motion predictions

How to select the optimal formulations for building blended physics-machine learning models for ship motions is not currently clear. This work compares and contrasts two approaches to this problem: (1) A black-box deep learning approach based on a new neural network architecture that can better handle varying wave conditions, and (2) a clear-box model based on updates to linear response amplitude operators via a Gaussian process regression. Both models are trained and evaluated on a dataset consisting of more than 15,000 30-minute-long motion observation windows from two research vessels at sea in the Atlantic and Pacific oceans. Three different hindcast weather services are used, including two models from the EU’s Copernicus system and NOAA’s WAVEWATCH III. The evaluation shows that a tradeoff exists between the formulations, with the black-box formulation offering higher accuracy and the cost of less transparency. The weather hindcast used has a small impact on the results, and the ability of both models to generalize predictions between near-sister ships is also encouraging for the practical application of these techniques.

A black box approach to fitting smooth models of mortality

Actuaries have long been interested in the forecasting of mortality for the purpose of the pricing and reserving of pensions and annuities. Most models of mortality in age and year of death, and often year of birth, are not identifiable so actuaries worried about what constraints should be used to give sensible estimates of the age and year of death parameters, and, if required, the year of birth parameters. These parameters were then forecast with an ARIMA model to give the required forecasts of mortality. A recent article showed that, while the fitted parameters were not identifiable, both the fitted and forecast mortalities were. This result holds if the age term is smoothed with P-splines. The present article deals with generalized linear models with a rank deficient regression matrix. We have two aims. First, we investigate the effect that different constraints have on the estimated regression coefficients. We show that it is possible to fit the model under different constraints in R without imposing any explicit constraints. R does all the necessary booking-keeping ‘under the bonnet’. The estimated regression coefficients under a particular set of constraints can then be recovered from the invariant fitted values. We have a black box approach to fitting the model subject to any set of constraints.

Accurate Structures and Spectroscopic Parameters of Guanine Tautomers in the Gas Phase by the Pisa Conventional and Explicitly Correlated Composite Schemes (PCS and PCS-F12).

A general strategy for the accurate computation of structural and spectroscopic properties of biomolecule building blocks in the gas phase is proposed and validated for tautomeric equilibria. The main features of the new model are the inclusion of core-valence correlation in geometry optimizations by a double hybrid functional and the systematic use of wave-function composite methods in conjunction with cc-pVnZ-F12 basis sets with separate extrapolation of MP2 and post-MP2 contributions. The resulting Pisa composite scheme employing conventional (PCS) or explicitly correlated (PCS-F12) approaches is applied to the challenging problem of guanine tautomers in the gas phase. The results are in remarkable agreement with the experimental structures, relative stabilities, and spectroscopic signatures of different tautomers. The accuracy of the results obtained at reasonable cost by means of black-box parameter-free approaches paves the way toward systematic investigations of other molecular bricks of life also by non-specialists.

Multi-Abstractive Neural Controller: An Efficient Hierarchical Control Architecture for Interactive Driving

As learning-based methods make their way from perception systems to planning/control stacks, robot control systems have started to enjoy the benefits that data-driven methods provide. Because control systems directly affect the motion of the robot, data-driven methods, especially black box approaches, need to be used with caution considering aspects such as stability and interpretability. In this paper, we describe a differentiable and hierarchical control architecture. The proposed representation, called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">multi-abstractive neural controller</i> , uses the input image to control the transitions within a novel discrete behavior planner (referred to as the visual automaton generative network, or <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">vAGN</i> ). The output of a vAGN controls the parameters of a set of dynamic movement primitives which provides the system controls. We train this neural controller with real-world driving data via behavior cloning and show improved explainability, sample efficiency, and similarity to human driving.

Flow Coefficient Determination in Catchment Based on Analysis of Temperature and Wind Speed Data Using the Fuzzy SMRGT Method

In engineering hydrology, calculating the flow coefficient is a crucial step. The flow coefficient calculation is necessary for directing the rational profiteering of water resources, improving the overall efficiency of water resource utilization, and minimizing the effect of catastrophic events. By precisely determining the flow coefficient, which is the most influential factor in flood flow, the current issues will be mitigated substantially. Various techniques are available in the existing literature for modelling flow coefficient. Most of them, however, rely on black-box approaches that are not generalizable. Therefore, this paper applied an intelligent model based on a fuzzy logic system called the Simple Membership Function and Fuzzy Rules Generation Technique (SMRGT). The new technique considers the physical cause-effect relationship and is intended to aid individuals who struggle to choose the number, form, and logic of membership functions and fuzzy rules in any fuzzy set. The study area’s temperature and wind speed data were incorporated into the SMRGT model’s input variables. The output was the flow coefficient. The prediction made by the model was validated against observational data. The comparison relies on numerous statistics and errors. The results indicated that the SMRGT model predicts the flow coefficient extraordinarily well and is an excellent method for generating membership functions and fuzzy rules.

Model predictive control of nonlinear processes using neural ordinary differential equation models

Neural Ordinary Differential Equation (NODE) is a recently proposed family of deep learning models that can perform a continuous approximation of a linear/nonlinear dynamic system using time-series data by integrating the neural network model with classical ordinary differential equation solvers. Modeling nonlinear dynamic processes using data has historically been a critical challenge in the field of chemical engineering research. With the development of computer science and neural network technology, recurrent neural networks (RNN) have become a popular black-box approach to accomplish this task and have been utilized to design model predictive control (MPC) systems. However, as a discrete-time approximation model, RNN requires strictly uniform step sequence data to operate, which makes it less robust to an irregular sampling scenario, such as missing data points during operation due to sensor failure or other types of random errors. This paper aims to develop a NODE model and construct an MPC based on this novel continuous-time neural network model. In this context, closed-loop stability is established and robustness to noise is addressed using a variety of data analysis techniques. An example of a chemical process is utilized to evaluate the performance of NODE-based MPC. Furthermore, the performance of the NODE-based MPC under Gaussian and non-Gaussian noise is investigated, and the subsampling method is found to be effective in building models for MPC that are suitable for handling the presence of non-Gaussian noise in the data.

Combining Prior Knowledge and Data for Robust Controller Design

We present a framework for systematically combining data of an unknown linear time-invariant system with prior knowledge on the system matrices or on the uncertainty for robust controller design. Our approach leads to linear matrix inequality (LMI) based feasibility criteria which guarantee stability and performance robustly for all closed-loop systems consistent with the prior knowledge and the available data. The design procedures rely on a combination of multipliers inferred via prior knowledge and learnt from measured data, where for the latter a novel and unifying disturbance description is employed. While large parts of the paper focus on linear systems and input-state measurements, we also provide extensions to robust output-feedback design based on noisy input-output data and against nonlinear uncertainties. We illustrate through numerical examples that our approach provides a flexible framework for simultaneously leveraging prior knowledge and data, thereby reducing conservatism and improving performance significantly if compared to black-box approaches to data-driven control.

Embedded Localized Molecular-Orbital Representations for Periodic Wave Functions.

To more straightforwardly provide local chemical-bonding reasoning in crystalline matter, we introduce a new approach to generate a real-space analogue of periodic electronic structures using "exact" top-down frozen-density embedding calculations. Based on the obtained real-space electronic structure, we then construct localized molecular orbitals and evidence that our technique compares favorably against the commonly used Wannier method, both in terms of numerical efficiency and details of chemical bonding. The new method has been implemented into the LOBSTER software package and designed as a black-box approach, digesting any periodic electronic structure from the currently supported codes, i.e., VASP, Quantum ESPRESSO, and ABINIT.

An Assessment of the Effectiveness of the Remedial Teaching Education Policy

The remedial teaching policy is a flagship education policy of the Greek Ministry of Education that aims to create a school of equal opportunities by providing additional support to students from disadvantaged social backgrounds. In this work we utilised a data set provided by the Ministry of Education, followed a black box approach and built on previous results in order to achieve the first ever evaluation, based on data, of the remedial teaching policy. Our findings indicate that remedial teaching is very effective in supporting very weak students, helping 70% of them achieve better academic performance and one out of three of them to sustain this enhanced academic performance in the future, long after they have stopped receiving remedial teaching. On the other hand, and contrary to what is widely believed, our results show that remedial teaching has the opposite impact to what it was designed for, as it is primarily the privileged students that receive the benefits. Consequently, in the way it is currently implemented, remedial teaching widens the gap between privileged and disadvantaged students rather than reduces it. The implications of the work are wide and far reaching, including the establishment of the need to revisit the way remedial teaching is implemented, the highlighting of the value in the data gathered by the Ministry of Education and the proof that individual educational policies can be objectively assessed despite being part of a complex system in which multiple education policies are implemented concurrently.

Fabrication of Polyamide6/Polyaniline as an Effective Nano-web Membrane for Removal of Cr (VI) from Water and a Black Box Approach in Modeling of Adsorption Process.

Chromium (Cr), as a highly toxic heavy metal ion, is still a severe environmental issue, although many research efforts have been put into its removal from water. Polyaniline (PANI), as a conductive polymer, demonstrated great capability in heavy metal adsorption due to its low cost, ease of synthesis, reversible redox behavior, and chemical stability. However, using PANI powder alone in heavy metal removal causes secondary pollution and aggregation in water. The PANI coating on a substrate could tackle this problem. In this study, polyaniline-coated polyamide6 (PA6/PANI) nano-web membrane was used for the removal of Cr(VI) in both adsorption and filtration-adsorption modes. The PA6/PANI nano-web membrane was fabricated via PA6 electrospinning followed by in-situ polymerization of the aniline monomer. The electrospinning condition of PA6 was optimized by the Taguchi method. The PA6/PANI nano-web membrane was characterized by FESEM, N2-adsorption/desorption, FT-IR, contact angle measurement, and tensile test. FT-IR and FESEM results demonstrated the successful synthesis of PA6/PANI nano-web and PANI homogeneous coating on PA6 nanofibers, respectively. The N2 adsorption/desorption results indicated that the pore volume of the PA6/PANI nano-web decreased by 39% compared to PA6 nanofibers. The tensile test and water contact angle studies showed that the coating of PANI on PA6 nanofibers improves the mechanical properties and hydrophilicity of PA6 by 10% and 25%, respectively. The application of PA6/PANI nano-web in the removal of Cr(VI) in batch and filtration modes exhibits excellent removal of 98.4 and 86.7%, respectively. A pseudo first order model well described the adsorption kinetics, and the adsorption isotherm was best fitted by the Langmuir model. A black box modeling approach based on artificial neural networks (ANN) was developed to predict the removal efficiency of the membrane. The superior performance of PA6/PANI in both adsorption and filtration-adsorption systems makes it a potential candidate for the removal of heavy metals from water on an industrial scale.

A Gradient-Descent Optimization Assisted Gray-Box Impedance Modeling of EV Chargers

Extracting an electric vehicle (EV) charger's input impedance with the analytical model (white-box approach) or the frequency sweep (black-box approach) is limited by the parameter confidentiality or the measurement noise, respectively. To overcome these challenges, a Gradient-Descent (GD) optimization based gray-box modeling approach is proposed. To start with, a sensitivity study on the analytical impedance model of an EV charger with a typical controller is carried out to identify the influential frequency range per controller and circuit parameter. On top of that, given an EV charger with unknown control and circuit information, a GD optimization based algorithm for multiple parameter estimation is designed to identify the unknown controller and circuit parameters based on the measured impedance, by assuming the EV charger is using the typical controller. Then, an analytical input impedance of the black-box EV charger can be obtained. Moreover, the low accuracy issue commonly encountered when estimating multiple parameters with GD optimization is mitigated with the proposed algorithm. Compared to pure frequency sweep, the proposed approach achieves a higher accuracy for the coupling impedance and a comparable accuracy for the diagonal impedance. The effectiveness of the proposed approach is validated by experimental results.

Simulating Realistic Continuous Glucose Monitor Time Series By Data Augmentation.

Simulated data are a powerful tool for research, enabling benchmarking of blood glucose (BG) forecasting and control algorithms. However, expert created models provide an unrealistic view of real-world performance, as they lack the features that make real data challenging, while black-box approaches such as generative adversarial networks do not enable systematic tests to diagnose model performance. To address this, we propose a method that learns missingness and error properties of continuous glucose monitor (CGM) data collected from people with type 1 diabetes (OpenAPS, OhioT1DM, RCT, and Racial-Disparity), and then augments simulated BG data with these properties. On the task of BG forecasting, we test how well our method brings performance closer to that of real CGM data compared with current simulation practices for missing data (random dropout) and error (Gaussian noise, CGM error model). Our methods had the smallest performance difference versus real data compared with random dropout and Gaussian noise when individually testing the effects of missing data and error on simulated BG in most cases. When combined, our approach was significantly better than Gaussian noise and random dropout for all data sets except OhioT1DM. Our error model significantly improved results on diverse data sets. We find a significant gap between BG forecasting performance on simulated and real data, and our method can be used to close this gap. This will enable researchers to rigorously test algorithms and provide realistic estimates of real-world performance without overfitting to real data or at the expense of data collection.