Scientific modeling often requires discovering mathematical relationships that generalize across multiple related datasets while capturing context-specific deviations. We present a multiview symbolic regression method that decomposes predictions into a shared core equation and view-specific residual corrections. Through systematic ablation studies, we identify domain-informed initialization as the critical success factor: seeding the evolutionary population with common scientific building blocks (polynomials, trigonometric functions, and rational expressions) improves performance by two orders of magnitude compared to random initialization. The multiview decomposition is optimized via NSGA-II with three objectives: prediction accuracy, core complexity, and residual complexity. Additional mechanisms, including elite protection and motif-based enhancement provide consistent but marginal gains. We evaluate our approach on synthetic benchmarks with known ground-truth decompositions and on oceanographic data modeling microbial metabolism across depth layers. On the ocean dataset, our method achieves 99% lower mean squared error than standard multiview symbolic regression, gradient boosting, and neural baselines across 50 prediction targets, while producing interpretable equations that reveal temperature and oxygen as dominant environmental drivers. Our results demonstrate that effective multiview symbolic regression depends more on appropriate initialization of the search space than on sophisticated search operators.

Business documents often contain substantial tabular and textual information with numerical values, requiring mathematical reasoning for effective document understanding. While Small Language Models (SLMs) still struggle at this task, tool-augmented multi-step agents perform better, at the cost of relying on closed-source or larger models, external data, or extensive prompt-engineering. This work introduces MATATA, a novel weakly supervised end-to-end approach to train multi-step reasoning language agents for document tabular applications. MATATA presents an annotation-free paradigm for each agent to enhance 3.8B/8B SLMs. During its two-stage training, MATATA uses the final outcome of the multi-step reasoning chain as weak supervision. This approach avoids having to individually supervise each intermediate agent in the reasoning chain. By employing an adaptive planner and shared tools across different datasets, MATATA shows robust performance. Experiments demonstrate that MATATA achieves state-of-the-art on FinQA, and on TAT-QA among reasoning methods based on open-source SLMs. Although being SLM-based, MATATA closely matches GPT-4-based frameworks on TabMWP. This novel weakly supervised approach enables training an end-to-end multi-step reasoning agent without intermediate supervision, supporting future developments of cost-effective powerful agentic systems.

Recently, new Neural Language Models pre-trained on a massive corpus of texts are available. These models encode statistical features of the languages through their parameters, creating better word vector representations that allow the training of neural networks with smaller sample sets. In this context, we investigate the application of these models to predict Item Response Theory parameters in multiple choice questions. More specifically, we apply our models for the Brazilian National High School Exam (ENEM) questions using the text of their statements and propose a novel optimization target for regression: Item Characteristic Curve. The architecture employed could predict the difficulty parameter b of the ENEM 2020 and 2021 items with a mean absolute error of 70 points. Calculating the IRT score in each knowledge area of the exam for a sample of 100,000 students, we obtained a mean absolute below 40 points for all knowledge areas. Considering only the top quartile, the exam’s main target of interest, the average error was less than 30 points for all areas, being the majority lower than 15 points. Such performance allows predicting parameters on newly created questions, composing mock tests for student training, and analyzing their performance with excellent precision, dispensing with the need for costly item calibration pre-test step.

The mining sector is a very relevant part of the Chilean economy, representing more than 14% of the country’s GDP and more than 50% of its exports. However, mining is also a high-risk activity where health, safety, and environmental aspects are fundamental concerns to take into account to render it viable in the longer term. The Chilean National Geology and Mining Service (SERNAGEOMIN, after its name in Spanish) is in charge of ensuring the safe operation of mines. On-site inspections are their main tool in order to detect issues, propose corrective measures, and track the compliance of those measures. Consequently, it is necessary to create inspection programs relying on a data-based decision-making strategy. This paper reports the work carried out in one of the most relevant dimensions of said strategy: predicting the mining worksites accident risk. That is, how likely it is a mining worksite to have accidents in the future. This risk is then used to create a priority ranking that is used to devise the inspection program. Estimating this risk at the government regulator level is particularly challenging as there is a very limited and biased data. Our main contribution is to apply a multi-task learning approach to train the risk prediction model in such a way that is able to overcome the constraints of the limited availability of data by fusing different sources. As part of this work, we also implemented a human-experience-based model that captures the procedures currently used by the current experts in charge of elaborating the inspection priority ranking. The mining worksites risk rankings built by model achieve a 121.2% NDCG performance improvement over the rankings based on the currently used experts’ model and outperforms the non-multi-task learning alternatives.

This work extends current collective intelligence evolutionary algorithms by incorporating a collective-based variation operator. As part of this work, the proposals are compared with state-of-the-art reference-point-based MOEAs: NSGA-II and RNSGA-II. Another primary objective of the work is to deal with a real-world multi-objective instance of the facility location problem. The experimental results validate the proposal. The new collective intelligence MOEA outperformed NSGA-II and R-NSGA-II for complex scenarios.

This work proposes a new algorithm that combines Iterated Local Search (ILS), Variable Neighborhood Search (VNS), Random Variable Neighborhood Descent (RVND) and a constructive initial search to solve the Traveling Car Renter Problem (CaRS). The CaRS a variation of the Traveling Salesman Problem where a customer wants to visit a set of cities using a rental car. The customer has a fixed amount of car options available to choose and when he returns the car, he will not be able to use it again. The traveling cost varies according to the car used, being the same value in the opposite direction. In addition to this cost, the customer also pays for a vehicle return. The objective is to minimize costs and fees. The results were compared to the state of the art and showed better performances.

This paper puts forward a new text to tensor representation that relies on information compression techniques to assign shorter codes to the most frequently used characters. This representation is language-independent with no need of pretraining and produces an encoding with no information loss. It provides an adequate description of the morphology of text, as it is able to represent prefixes, declensions, and inflections with similar vectors and are able to represent even unseen words on the training dataset. Similarly, as it is compact yet sparse, is ideal for speed up training times using tensor processing libraries. As part of this paper, we show that this technique is especially effective when coupled with convolutional neural networks (CNNs) for text classification at character-level. We apply two variants of CNN coupled with it. Experimental results show that it drastically reduces the number of parameters to be optimized, resulting in competitive classification accuracy values in only a fraction of the time spent by one-hot encoding representations, thus enabling training in commodity hardware.

One of the principles of evolutionary multi-objective optimization is the conjoint optimization of the objective functions. However, in some cases, some of the objectives are easier to attain than others. This causes the population to lose diversity at a high rate and stagnate in early stages of the evolution. This paper presents the progressive addition of objectives (PAO) heuristic. PAO gradually adds objectives to a given problem relying on a perceived measure of complexity. This diversity loss phenomenon caused by the nature of a given objective has been observed when applying the Voronoi diagram-based evolutionary algorithm (VorEAl) in anomaly detection problems. Consequently, PAO has been first directed to address that issue. The experimental studies carried out show that the PAO heuristic manages to yield better results than the direct use of VorEAl on a group of test problems.

Selection methods are a key component of all multi-objective and, consequently, many-objective optimisation evolutionary algorithms. They must perform two main tasks simultaneously. First of all, they must select individuals that are as close as possible to the Pareto optimal front (convergence). Second, but not less important, they must help the evolutionary approach to provide a diverse population. In this paper, we carry out a comprehensive analysis of state-of-the-art selection methods with different features aimed to determine the impact that this component has on the diversity preserved by well-known multi-objective optimisers when dealing with many-objective problems. The algorithms considered herein, which incorporate Pareto-based and indicator-based selection schemes, are analysed through their application to the Walking Fish Group (WFG) test suite taking into account an increasing number of objective functions. Algorithmic approaches are assessed via a set of performance indicators specifically proposed for measuring the diversity of a solution set, such as the Diversity Measure and the Diversity Comparison Indicator. Hypervolume, which measures convergence in addition to diversity, is also used for comparison purposes. The experimental evaluation points out that the reference-point-based selection scheme of the Non-dominated Sorting Genetic Algorithm III (NSGA-III) and a modified version of the Non-dominated Sorting Genetic Algorithm II (NSGA-II), where the crowding distance is replaced by the Euclidean distance, yield the best results.

It can be argued that in order to produce a sub-stantial improvement in multi-objective estimation of distribution algorithms it is necessary to focus on a particular group of issues, in particular, on the weaknesses derived from multi-objective fitness assignment and selection methods, the incorrect treatment of relevant but isolated (precursor) individuals; the loss of population diversity, and the use of ‘general purpose’ modeling algorithms without taking note of the particular requirements of the task. In this work we introduce the S-Metric Selection Estimation of Distribution Algorithm based on Multivariate Extension of Copulas (SMS-EDA-MEC). SMS-EDA-MEC was devised with the intention of dealing with those issues in mind. It builds the population model relying on the comprehensive Clayton's copula and incorporates methods for automatic population restarting and for priming precursor individuals. The experimental studies presented show that SMS-EDA-MEC yields better results than current and ‘traditional’ approaches.

The multi-objective neural EDA (MONEDA) was proposed with the aim of overcoming some difficulties of current MOEDAs. MONEDA has been shown to yield relevant results when confronted with complex problems. Furthermore, its performance has been shown to adequately adapt to problems with many objectives. Nevertheless, one key issue remains to be studied: MONEDA scalability with regard to the number of decision variables. In this paper has a two-fold purpose. On one hand we propose a modification of MONEDA that incorporates an indicator-based selection mechanism based on the HypE algorithm, while, on the other, we assess the indicator-based MONEDA when solving some complex two-objective problems, in particular problems UF1 to UF7 of the CEC 2009 MOP competition, configured with a progressively-increasing number of decision variables.

Soft computing methods, and Multi-Objective Evolutionary Algorithms (MOEAs) in particular, lack a general convergence criterion which prevents these algorithms from detecting the generation where further evolution will provide little improvements (or none at all) over the current solution, making them waste computational resources. This paper presents the Least Squares Stopping Criterion (LSSC), an easily configurable and implementable, robust and efficient stopping criterion, based on simple statistical parameters and residue analysis, which tries to introduce as few setup parameters as possible, being them always related to the MOEAs research field rather than the techniques applied by the criterion.

In this work we present a novel progress indicator, called fitness homogeneity indicator (FHI). This indicator improves the other previously discussed indicators as it takes into account all possible processes taking place in the population while not requiring an intensive computation as it relies on the fitness values calculated for the individuals. It is also capable of equally detecting success and failure scenarios, hopefully making an early detection of the second case.

In this work we analyze the model-building issue and the requirements it imposes on the learning paradigm being used. We argue that error-based learning, the class of learning most commonly used in MOEDAs, is responsible for current MOEDA underachievement. We present ART as a viable alternative and present a novel algorithm called multi-objective ART-based EDA (MARTEDA) that uses a Gaussian ART neural network for model-building and an hypervolume based selector as described for the HypE algorithm. We experimentally show that thanks to MARTEDA's novel model-building approach and an indicator-based population ranking the algorithm it is able to outperform similar MOEDAs and MOEAs.

The need for a stopping criterion in MOEA's is a repeatedly mentioned matter in the domain of MOOP's, even though it is usually left aside as secondary, while stopping criteria are still usually based on an a-priori chosen number of maximum iterations. In this paper we want to present a stopping criterion for MOEA's based on three different indicators already present in the community. These indicators, some of which were originally designed for solution quality measuring (as a function of the distance to the optimal Pareto front), will be processed so they can be applied as part of a global criterion, based on estimation theory to achieve a cumulative evidence measure to be used in the stopping decision (by means of a Kalman filter). The implications of this cumulative evidence are analyzed, to get a problem and algorithm independent stopping criterion (for each individual indicator). Finally, the stopping criterion is presented from a data fusion perspective, using the different individual indicators' stopping criteria together, in order to get a final global stopping criterion.

The multi-objective optimization neural estimation of distribution algorithm (MONEDA) was devised with the purpose of dealing with the model-building issues of MOEDAs and, therefore address their scalability.

In order to comprehend the advantages and short-comings of each model-building algorithm they should be tested under similar conditions and isolated from the MOEDA it takes part of. In this work we will assess some of the main machine learning algorithms used or suitable for model-building in a controlled environment and under equal conditions. They are analyzed in terms of solution accuracy and computational complexity. To the best of our knowledge a study like this has not been put forward before and it is essential for the understanding of the nature of the model-building problem of MOEDAs and how they should be improved to achieve a quantum leap in their problem solving capacity.

In this paper we explore the model-building issue of multiobjective optimization estimation of distribution algorithms. We argue that model-building has some characteristics that differentiate it from other machine learning tasks. A novel algorithm called multiobjective neural estimation of distribution algorithm (MONEDA) is proposed to meet those characteristics. This algorithm uses a custom version of the growing neural gas (GNG) network specially meant for the model-building task. As part of this work, MONEDA is assessed with regard to other classical and state-of-the-art evolutionary multiobjective optimizers when solving some community accepted test problems.

In this work we present a novel and efficient algorithm independent stopping criterion, called the MGBM criterion,suitable for Multi-objective Optimization Evolutionary Algorithms(MOEAs).The criterion, after each iteration of the optimization algorithm, gathers evidence of the improvement of the solutions obtained so far. A global (execution wise) evidence accumulation process inspired by recursive Bayesian estimation decides when the optimization should be stopped. Evidence is collected using a novel relative improvement measure constructed on top of the Pareto dominance relations. The evidence gathered after each iteration is accumulated and updated following a rule based on a simplified version of a discrete Kalman filter.Our criterion is particularly useful in complex and/or high-dimensional problems where the traditional procedure of stopping after a predefined amount of iterations cannot beused and the waste of computational resources can induceto a detriment of the quality of the results.Although the criterion discussed here is meant for MOEAs,it can be easily adapted to other soft computing or numerical methods by substituting the local improvement metric witha suitable one.

In this work we propose a hybrid neural architecture named VaBiARRT based on the adaptive resonance theory that relies on representational redescription to archive a high degree of generalization and code compactness. Knowledge is autonomously structured as a hierarchical topology of fuzzy input classes and how these classes are related with the outputs. A two-way abstraction/particularization process takes place in order to rewrite the established relations to make them as abstract as possible without loosing accuracy. The internal handling of the representation of knowledge can be interpreted as a standard pattern matching variable binding process. Besides providing a formal description of VaBiARRT we also solve a sample problem related to the representational redescription hypothesis in order to study the knowledge redescription process and the performance of the network.