A Mathematical Programming (MP) model is a common representation for virtually all types of real-world entities, e.g., production lines, delivery schedules, personnel assignment. The MP model consists of variables that correspond to, e.g., production volume, constraints that represent the relationships between the variables, e.g., production conditions, and an objective function that corresponds to the outcome, e.g., production cost. A feasible solution is a vector of values of variables that satisfies all constraints. The feasible solution that minimizes the objective function is the optimal solution. The use of MP models is fully automated thanks to the solvers — the tools that produce solutions, either feasible or optimal. However, building the correct MP model requires intense training and expertise, and often turns out to be time-consuming and error-prone. The errors in MP models are difficult to spot and often remain unidentified until the optimal solution turns out to be inapplicable in practice, resulting in many iterations of modeling, conformance checking, and model enhancement. MP models have many advantages over many other representations, e.g., interpretable structure offering explainable decisions, executable semantics, and mature solver tools. The project aim is to help human experts by developing new MP Model Mining algorithms. Our contribution is divided into four areas: discovery, conformance checking, enhancement, and verification in real-world settings. We use easy-to-provide data, such as exemplary solutions recorded by computers managing the modeled entity, available quantities, e.g., parameters, sets, variables, and/or other background knowledge information. Novel discovery algorithms will build from the data the MP models that maximize three criteria: fitness, precision, and generalization. Fitness evaluates how well the MP model embraces the data, precision assesses how tight the MP model is, and generalization measures how well the MP model describes different instances of the modeled entity. Novel conformance checking algorithms will calculate these measures given the existing MP model and the data. Novel enhancement algorithms will identify deficiencies in the existing MP model and propose fixes. We will develop algorithms that employ high-level modeling languages, e.g., AMPL and ZIMPL. We will also propose new measures for fitness, precision, and generalization, as the contemporary Artificial Intelligence measures are largely inappropriate to MP models, e.g., award the models equally for all examples and ignore syntax while for most MP models concise representation and tight constraints are more important. We will also seek for new algorithms that propose fixes in MP models driven by, e.g., irreducible inconsistent subsystem and counterexample data. To date, most of the work in the field has been done in the context of the discovery problem. However, the existing algorithms are immature and are not ready to be used in large-scale real-world problems. They suffer from the curse of dimensionality for even 6-8 variables, do not handle well one-class data, and/or have unacceptable computation complexity. Our ultimate goal is to overcome these challenges. The only way to reliably validate the developed algorithms is their use to build MP models for complex real-world Operations Research (OR) problems. To this aim, we will investigate current and past OR challenges, e.g., ROADEF and PACE challenges. They consist of problems of high-importance in industry and data donated by international companies. Exemplary past problems in these challenges are truck loading by Renault, maintenance planning by RTE, cutting by Saint-Gobain, and inventory routing by Air Liquide. We will use the proposed algorithms to develop proof-of-concept MP models ready to solve these OR problems. Our contributions will make foundations for better modeling and optimization tools for almost every sector of economy. Note that solving the problems originating in industry is not the project goal per se but only a side effect of basic research on algorithms for MP model mining.
The aim of the Produktoskop project is to create an innovative system that searches and reports on products about which there are doubts concerning their quality, with particular attention to cases of Dual Quality practices used by entrepreneurs, on the basis of analysis of consumer opinions posted on the Internet. In order to build the system, artificial intelligence methods will be developed and used, which is in line with the main objective of the Infostrateg programme.
This project develops neurosymbolic algorithms for program synthesis – automatically generating code from specifications – by combining symbolic reasoning with deep learning. Focusing on iterative methods, the approach enables systems to learn and refine partial solutions, improving efficiency and generalisation across tasks. The research explores graph-based representations, suitable neural architectures, and learning paradigms to guide synthesis. Theoretical models will be implemented, tested, and evaluated against state-of-the-art methods using formal metrics. The outcomes aim to advance both program synthesis and broader machine learning applications that involve complex, combinatorial reasoning.
As part of this project, new explainable AI models will be developed for predicting the properties of chemical compounds, generative models proposing new chemical substrates for battery creation, and a tool application supporting the work of chemists using the aforementioned AI models. The project will also result in an open database of experimentally (chemically and physically) tested new battery substrates.
The project concerns the issues of machine learning and artificial intelligence, and more specifically, learning predictive models from evolving data streams with elements of explaining the operation of these models (the so-called XAI paradigm for explaining the operation of machine learning models) and changes occurring in the data themselves.
The project focuses on creating algorithms for explainable decision-making systems and measuring fairness. Theoretical properties of existing fairness metrics are being investigated, and new metrics are being designed that take into account data imbalance. The project also develops methods for visualizing multi-criteria decision-making systems, which will make it easier to explain the decisions they make. We will use the results of our research in practical applications and make them available to the general public in the form of explainable multi-criteria decision dashboards and libraries for programmers
The project deals with digital logic design using FPGA platforms, hardware design for embedded systems using Intel FPGA architecture, and RTL to GDS synthesis. In addition, the project's topics hook into RTL implementation, modern heterogeneous architectures, synchronous circuits, simulation and debug, static timing analysis, design constraints, memory subsystems, System of Chips (SoCs) and System in Package (Sip).
Natwork for developing the scientific foundations for Trustworthy AI through the integration of learning, optimisation and reasoning
The aim of this project will be development of method for solving task of knowledge retrieval from LinkedmData. This will enable Linked Data’s more efficient utilization by a person looking for the answer for his/her query, as well as by machine learning methods
The project aims to advance the field of program synthesis by developing algorithms that automatically discover and utilize domain knowledge to enhance the efficiency of searching the program space. Program synthesis, the process of generating a program that meets a given specification, is inherently challenging due to its NP-hard nature. While heuristic algorithms can provide practical solutions, they are typically independent of the specific problem and unable to leverage its unique characteristics. The project focuses on utilizing machine learning techniques to identify useful properties of program synthesis problems, which can then be used to prioritize the search process and reduce the time required to solve these problems.
A business process, understood as a set of activities in dependency relation leading to the achievement of a business goal, is an essential part of every enterprise. Effective management of business processes in medium and large enterprises is impossible without the support of dedicated software. The aim of the project is to develop ProcessM, a fully-functional software product for modeling, monitoring, analysis and optimization of business processes. ProcessM is not an Enterprise Resource Planning (ERP) software, but an artificial intelligence tool supporting business process management. ProcessM seamlessly integrates with ERP tools and any other data sources thanks to the use of machine learning methods that support extraction, transformation and loading (ETL) of data. ProcessM models the business process in online mode based on the incoming in real-time events from the ERP system, verifies the conformance of the process with the model, classifies detected deviations from the model into errors and concept drift, reports deviations together with a root cause analysis and proposes changes to the process to optimize its performance. ProcessM is a webservice that works in a continuous mode and does the above tasks unattended. Detected deviations in the operation of the process are immediately reported to the user. The web-service of ProcessM is compatible with Windows, Linux and MacOS, and the client of ProcessM works in a web-browser running on virtually any operating system.
Research project objectives The project is aimed at reducing costs of building Mathematical Programming (MP) models for business processes by automating this task. The common practice is to prepare MP models manually using expert knowledge, however this is laborious task that requires deep insight into both, the business process and the modeling techniques. We propose algorithms that automate model building using examples of the process states acquired using monitoring process execution. We decompose the problem of MP model synthesis into subproblems of constraint synthesis and objective function synthesis and tackle them separately using methods from Machine Learning, Computational Intelligence and Operational Research. To handle the constraint synthesis problem we propose algorithms based on one-class classifiers (e.g., Support Vector Data Description, POSC4.5), Genetic Programming, local search heuristics (e.g., Simulated Annealing, Tabu Search) and Linear Programming. To solve the objective function synthesis problem, we employ least-squares regression, including Gauss-Newton algorithm for non-linear regression, and symbolic regression using Genetic Programming. All the developed algorithms are analyzed theoretically and compared in computational experiments on properties of them and the MP models that they synthesize. For objective assessment, we use benchmark problems of controlled complexity as well as problem instances of real-world processes. Research project methodology We formulate three research tasks, related to constraint synthesis, objective function synthesis and evaluation of the developed algorithms on modeling of real-world processes. In the first two tasks we use synthetic benchmarks formulated as MP models and sampled to acquire the desired number of training examples. These examples are used by the developed algorithms to synthesize MP models. We assess both the algorithms and the MP models that they produce. The former ones are assessed on properties like computational complexity, scaling with problem size and mean square training error, and the latter ones on syntactic and semantic measures of fidelity to benchmark MP models, particularly on angles between constraints, test-set accuracy, Jaccard index of feasible regions etc. In the third task, we verify applicability of the developed algorithms in real-world scenarios. We use data acquired by monitoring various processes to synthesize MP models for them. For this scope, we use public data sets from UCI Machine Learning Repository on e.g., production of energy at a power plant and production of cement mixture at a concrete plant. For a synthesized MP model, we verify its conformance with available domain knowledge of the modeled process. We apply the MP model to process simulation under different working conditions, i.e., with some variables set to specific values, and analyze how these settings influence MP model’s feasible region and objective function. We also apply the MP model to process optimization and verify feasibility of the optimal solution in reality and improvement w.r.t. the current process outcome. Significance of quantitative differences between properties of algorithms and MP models is verified statistically, particularly using ANOVA and Friedman’s test. Expected impact of the research project on the development of science, civilization and society Nowadays, most of MP models are created manually by domain experts offering specialized consulting services. Only a limited number of organizations afford to employ these services due to significant costs, however process optimization using MP models would bring them substantial profits and savings that increase their advantage in competitive business environment. We automate model building and expect that the developed algorithms decrease amount of expert knowledge needed to build correct MP models for real-world processes, reducing thus human-related effort and cost of modeling task. The fidelity of the MP model to reality is crucial in optimization, where the optimal solutions usually lie at the boundary of feasible region determined by the constraints, thus we expect that the synthesized MP models reflect reality to the extent that allows their adoption to optimize, simulate and explain process behavior. Achievement of that goal would lead to availability of modeling and optimization services to a broader range of entities than the contemporary methods. The improved efficiency of processes should directly contribute to increase of income and competitiveness of the organizations employing these processes and in the effect to increase of gross domestic product and wealth of society.
The aim of the project is to develop an innovative medical diagnosis support system based on artificial intelligence. The scope of research and development under the project concerns the design and development of IT solutions to support the diagnosis of diseases (skeletal trauma and malignant lung tumors) based on artificial intelligence, extended inference systems (machine learning) and systems based on computer simulations at different levels of complexity.
The goal of the project is to design and implement a toolbox of explanatory modeling for Apache Spark and to make it available to entities from business and academia. Explanatory modeling leverages the techniques of machine learning for automatic acquisition of transparent models from data produced by arbitrary processes, including business operations, product development/prototyping, scientific experimentation, etc., for making predictions, performing classification, clustering and other activities related to data analytics. Transparency is achieved by representing models as symbolic (algebraic or logic) expressions and rules of the form “if ... then ...”. Such models, in addition to serving as typical predictors (classifiers, regressors, etc.), offer insight into the nature of the process in question and facilitate various forms of explanation: explanation of predictions being made, their causes, interactions between parameters that characterize the process, etc. These features, though essential in many application areas, are largely absent in the mainstream machine learning methods like neural networks, random forests or support vector machines.
As part of the project, the CheckMyBlob system (https://checkmyblob.bioreproducibility.org/) was developed for automatically detecting, predicting, and validating ligands within the crystal structures of biological macromolecules. The system identifies potential ligand binding sites and uses machine learning to suggest which molecules might fit into a specific region of the electron density map. It can also validate previously modeled structures, including those deposited in the PDB (Protein Data Bank). Additionally, the project developed the PQ1 metric, which provides a straightforward way to assess the quality (in terms of fit to crystallographic data) of protein and nucleic acid structures, , and a neural network that recognizes artifacts in diffraction images from X-ray crystallography experiments.
Within the project, the AUE (Accuracy Updated Ensemble) algorithm was developed for classifying block-processed data streams. The developed algorithm maintains high classification accuracy under various types of changes occurring in the streams, such as sudden, gradual, incremental, or recurring changes in class definitions. Based on experiments comparing AUE with existing stream classifiers, it was shown that the proposed algorithm achieves, on average, the highest classification accuracy while maintaining low memory requirements and short processing time. The project also demonstrated the possibility of transferring solution elements existing in block classifiers to incrementally operating methods, and proposed and evaluated strategies for adapting block algorithms to incremental environments. Based on this analysis, the OAUE (Online Accuracy Updated Ensemble) algorithm was proposed, which incrementally trains and evaluates base classifiers according to the best response strategies to sudden and gradual types of changes. The created algorithms have become part of the MOA data stream analysis system (https://moa.cms.waikato.ac.nz/).
A new research method has been proposed for use in Earth sciences, particularly in the fields of paleontology, paleoecology, paleoceanography, and theoretical morphology. The primary goal was to construct a new algorithmic environment for testing and simulating the principles of evolution and their complex consequences for organisms under specific dynamic habitat conditions, varying over geological time and space. The model organism was foraminifera – single-celled eukaryotic organisms that densely inhabit marine pelagic and benthic zones and present an exceptionally complete fossil record from the entire Phanerozoic.
The aim of the project is to develop methods for producing an integrated actuator and diagnostics of soft actuators implemented using dielectric electroactivated polymers (DEAP) and magnetorheological elastomers (MRE). DEAP and MRE materials belong to the group of intelligent materials. DEAP actuators are characterized by high energy efficiency, flexibility and scalability. They are used in valves and pumps. Dielectric-electroactive polymer technology is able to meet specific requirements better than commonly used technology, e.g. electromagnets, but has strength and stroke restrictions.
The project, carried out in cooperation with PSNC and the Wasilewicz Eye Clinic, is devoted to the diagnosis of glaucoma that can be caused by increased intraocular pressure (IOP). Unlike diagnostic methods based on optical coherence tomography (OCT), the project is assuming analysis of Sensimed Triggerfish contact lens sensor data supplemented with additional clinical measurements. The aim of the project is to develop ML models for diagnosing the disease and to implement tools supporting the work of ophthalmologists.
The project consisted of the implementation of Edge computing technology at Poznan University of Technology and the preparation of training materials for second-level students pursuing studies at the Faculty of Computer Science and Telecommunications.
The project consisted of the implementation of embedded systems Intel technology at Poznan University of Technology and the preparation of training materials for second-level students pursuing studies at the Faculty of Computer Science and Telecommunications.