Marko Đurasević, University of Zagreb Faculty of Electrical Engineering and Computing

Website
Combinatorial optimization problems are encountered in many real-world situations and scenarios. Often, these problems are NP-hard, meaning that they cannot be solved optimally, but rather must be solved using various heuristic methods, most commonly improvement-based metaheuristics. However, for many real-world cases these improvement-based methods cannot be applied, since problems are stochastic or dynamic, and as such all the information about the problem is not available in advance, or the problems are simply too large to efficiently search the solution space, even heuristically. In those cases, constructive based heuristics usually represent the method of choice for solving such problems. Constructive based heuristics construct the solution step by step, only by determining the next decision that needs to be performed. Therefore, they can easily adapt to changes in the problem. The main issue with constructive heuristics is that they are difficult to design manually for the various problems that are encountered. For that reason, researchers have turned to the application of hyper-heuristics methods, i.e., methods that can be used not to solve the problem by themselves, but rather to generate a heuristic that can be used to solve any new problem of the type for which it was developed. Although many hyper-heuristic methods have been proposed in the literature, genetic programming has become the most popular method for that purpose. Over the years genetic programming has been used to generate heuristics for various combinatorial optimization problems, including various scheduling, routing and similar problems. Especially in recent years the research in this topic has been gaining on additional attention from the research community, with many different research directions being investigated, thus outlining the importance of this topic.

The goal of this tutorial is to provide a brief introduction into the topic of automated heuristic design with the use of generic programming. For that purpose, the tutorial will introduce genetic programming as an evolutionary computation method. After that, the application of genetic programming to the automated design of heuristics will be described, i.e., this part focuses on how genetic programming is used as a hyper-heuristic method. Here it will be outlined how the method needs to be adapted and how this problem is tackled, i.e., what steps are required to successfully apply genetic programming to generate a heuristic for a chosen combinatorial optimization problem. All of this will be done using a simple combinatorial optimization problem as a showcase. The tutorial will also present several applications of genetic programming to generate heuristics for various combinatorial optimization problems from the literature to outline the wide range applicability of the method, but also to emphasize differences and challenges encountered for various problems. The tutorial will cover some recent developments in the automatic generation of dispatching rules, as well as outline several new research directions in this field, such as multi-objective heuristic generation, application of ensemble learning methods, construction of surrogate models, etc. Finally, the current limitations and open issues of the approach will be outlined. The tutorial will help the interested researchers to acquire a good overview of this emerging and interesting research area, as well as the key ideas and challenges for future studies. Thus, the audience should gain knowledge that should help them successfully apply genetic programming to a given combinatorial optimization problem.

Nelishia Pillay, Department of Computer Science, University of Pretoria, South Africa
Thambo Nyathi, Department of Computer Science, University of Pretoria, South Africa

Website
Various machine learning, search and image processing techniques have been used to successfully predict diseases. The manual design of these approaches usually requires a number of design decisions and can be time-consuming and laborious. Furthermore, more recently prediction from more than one type of data, i.e., multimodal learning, may be necessary to predict diseases. A further challenge with using computational intelligence for disease prediction is imbalanced data. This tutorial focuses on the use of hyper-heuristics and evolutionary algorithms for the automated design of computational intelligence approaches for disease prediction. The tutorial firstly presents an overview of hyper-heuristics and evolutionary algorithms and how these techniques can be used to automate the design of neural networks, genetic programming and image processing techniques for disease prediction. The tutorial will then examine various case studies for disease prediction including the automated design of neural networks, genetic programming and image processing techniques as well as multimodal approaches and approaches catering for data imbalance. The case studies will be presented interactively using a tool developed for the automated design of computational intelligence techniques for disease prediction.

Prof. Chrystopher L. Nehaniv, University of Waterloo, Canada

Website
Cellular automata (CAs) are a widely applied model of massively parallel computation based on local neighborhoods and updates introduced by John von Neumann and Stanislaw Ulam. The tutorial introduces the concepts of cellular automata and examples, and overviews basic results, such as for the Game of Life (which is computationally universal and shows emergent properties), and variations of the cellular automata concept, including random boolean networks, synchronous and asynchronous automata networks, and discrete dynamical systems with external inputs. Von Neumann, one of the grandfathers, of Artificial Life and Computational Intelligence, also used CAs as a formal tool to study the logic of life and complexity: in particular, How is self-reproduction is possible, i.e., How is it possible for a mechanistic system to reproduce itself? How is it possible for something to produce something as or more complex than itself? Prior to the discovery of the structure of DNA and its relation to these questions, von Neumann gave several different solutions. It turns out that some correspond to life as we now know it, and others perhaps to life as it could be. We survey his solutions to these problems, and discuss progress on this question since then on self-reproducing systems. Also, we survey open problems for Artificial Life and Computational Intelligence research that go beyond the state of the art in the synthesis of self-reproducing systems, and offer challenges for researchers entering the field including those related to the concepts of individuality, robustness, evolution and self-production (autopoiesis).

Asst. Prof. Dr. Neyre Tekbiyik Ersoy, Energy Systems Engineering, Cyprus International University, Turkiye

Website
Smart grid is the modernized version of the power grid. It includes two-way communication systems and enables the integration of various technologies that will improve the grid's efficiency, reliability, sustainability, and security. Smart grid allows monitoring everything from power plants to consumer preferences and thus having more information about what is going on in the grid at every instant. In a smart grid, consumers may have an active role depending on demand response applications and utilization of distributed energy resources. This allows the countries to improve their energy security, by using more renewables. Considering the increasingly visible effects of global warming, such as; more frequent wildfires, longer droughts, and extreme weather events, it is evident that renewable energy resources should contribute more to the energy mix in the future. Smart grid may be the key for accomplishing that. The smart grid utilizes distributed computing and communications in order to deliver real-time information and enable the near-instantaneous balance of supply and demand at the device level. This way, the uncertainty and intermittency of renewables do not cause a burden on the grid, and their use can be increased. Hence, it can be said that if more renewables are to be integrated into the grid, then the grid should be updated towards smart grid and if smart grid exists, then more renewables can be integrated into the grid. Hence, this tutorial is aimed to explain the link between smart grid and the renewable energy integration to the grid.

Kai Olav Ellefsen, University of Oslo
Kyrre Glette, University of Oslo
Ege de Bruin, University of Oslo
Frank Veenstra, Nord University

Website
The evolution of robot bodies and brains allows researchers to investigate which building blocks are interesting for evolving Artificial Life, and how controllers and morphologies can be shaped together for automated robot design. This tutorial aims to introduce evolution of robot body and control, and some of the key challenges one faces when doing experiments in Evolutionary Robotics. These include finding good ways to represent robots (genotypic encodings), challenges related to co-optimizing morphology and control, how environments shape body and control, and selecting the right physical substrate for evolving robots. After introducing these challenges and showing relevant examples from our own and other labs' research, we will present a demo of how to run Evolutionary Robotics experiments in practice, with the Unity-based Evolving Modular Robots framework.

Michael Hellwig, Steffen Finck and Hans-Georg Beyer, Vorarlberg University of Applied Sciences, Austria

Website
Many real-world optimization problems come with restrictions originating from limited resources, physical dependencies, etc. Such restrictions (or constraints) narrow the search domain and may affect the structure of the optimization problem. In the field of Evolutionary Computation, constrained optimization problems are often tackled by penalty approaches in a first attempt. Yet, the addition of penalty functions affects the characteristics of the optimization problem. This can have a significant impact on the search behavior and lead, for example, to premature convergence in local attractor regions or to a steady retraction from the boundary of the feasible region of the search space on which the optimizer might be located. To alleviate these issues, different adaptive penalty approaches and algorithmic adjustments have been developed. The latter include repair mechanisms, which guarantee a search within the feasible region, or specially designed strategy parameter adaptation mechanisms to meet the characteristics of the restricted search space.

The tutorial focuses on state-of-the-art Evolution Strategies (ES) for continuous constrained single-objective optimization problems. Variants of the Matrix Adaptation Evolution Strategy proved successful in several benchmark competitions in the context of the IEEE CEC and the ACM GECCO conferences. The tutorial will elaborate on the suitability of ES in constrained environments and provide an overview of common benchmark frameworks in the domain. An emphasis is placed on design principles for successful constraint-handling operators. In this respect, we discuss the advantages and disadvantages of different interior and exterior point approaches on selected benchmark instances. Finally, the applicability of ES is demonstrated on selected real-world problem instances and potential lines of future research are discussed.

The tutorial is aimed equally at researchers and practitioners, who want to familiarize themselves with the mechanisms of ES for constraint handling, and who are looking for potentially suitable algorithms for constrained real-world problems. Python Jupyter Notebooks are used to create an interactive learning environment.

Samuel Yen-Chi Chen, Wells Fargo, New York
Joongheon Kim, Korea University, Korea

Website
This tutorial offers a hands-on, immersive introduction to the dynamic intersection of quantum machine learning (QML) and computational intelligence. It begins with a fundamental overview of quantum information science (QIS), covering key components such as qubits, quantum gates, measurements, and entanglement. From there, the session advances to core QML principles, exploring parameterized quantum circuits, data encoding techniques, and quantum circuit design methodologies. Participants will dive into a range of QML models, including quantum support vector machines (QSVM), quantum neural networks (QNN), and quantum convolutional neural networks (QCNN). The tutorial also pushes into the frontier of QML with advanced models like quantum recurrent neural networks (QRNN) and quantum reinforcement learning (QRL). Notably, it highlights the relevance of quantum models in computational intelligence, such as in multi-agent quantum reinforcement learning, illustrating their applicability in real-world, multi-agent scenarios. Through practical programming examples and demonstrations using open-source quantum simulators, attendees will gain concrete insights into how QML can enhance computational intelligence tasks. Designed for beginners, the tutorial provides a clear path for those eager to integrate quantum techniques into their research. It also offers guidance on advanced learning resources, software packages, and frameworks to extend their exploration beyond the session.

Prof. Wenwu Wang, University of Surrey, UK

Cross-modal generation between audio and language has become a significant area of research in both audio signal processing and natural language processing. Audio-to-language generation, often referred to as automated audio captioning, seeks to create meaningful textual descriptions of audio clips. This technology can assist hearing-impaired individuals in interpreting environmental sounds, enhance multimedia content retrieval, and support sound analysis for security surveillance. On the other hand, language-to-audio generation focuses on producing sound from a textual description, creating audio that matches the specified language prompt. Applications of this include sound design for films, games, virtual reality/metaverse environments, digital media, and tools for helping visually impaired users interpret text through sound. This tutorial will present language-audio models designed for mapping and aligning audio with textual data, their applications in cross modal language-audio learning, the creation of language-audio datasets, and potential future directions in this field. We will showcase recent advancements in large language-audio generation models, such as AudioLDM, AudioLDM2, and WavJourney for audio generation and storytelling, WavCraft for content creation and editing, and ACTUAL for audio captioning, along with captioning systems for generating diverse captions from audio clips. In addition, we will demonstrate language-audio learning applied to other related tasks, including AudioSep for audio source separation, SemantiCodec for audio coding, and APT-LLMs for audio reasoning. Finally, we will introduce datasets, such as WavCaps, Sound-VECaps, and AudioSetCaps, that can be used for training and evaluating large language-audio models.

Stefano Mintchev, ETH Zurich
Karine Miras, Vrije Universiteit Amsterdam
Mihir Kulkarni and Kostas Alexis, Norwegian University of Science and Technology
Marcelo Jacinto, Institute for Systems and Robotics
Leonard Bauersfeld and Davide Scaramuzza, University of Zurich

Website
Aerial robot simulation tools have experienced a leap in capacity and capability and currently facilitate robust learning of complex control policies through data. Exploiting progress in high-fidelity physics simulation, improved graphics, and the strengths of GPU-accelerated massively parallelized computing, modern simulators enable robots to collect millions of ‘experiences’ in a short time, thus allowing scalable learning and robust transfer to reality. Within this framework, there is a rising interest in simulators that allow the modeling of arbitrary aerial robot embodiments and simultaneously offer the capacity to generatively create environments either by using neural methods or parametric setting of reconfigurable scenes. The purpose is to automate the process of the computational design of novel aerial robots and their embodied AI solutions through efficient cycles of simulation, learning and adaptation.

Per Kristian Lehre, School of Computer Science, University of Birmingham, UK

A rich theory on runtime analysis (also called time-complexity analysis) of EAs has been developed over the last 20 years. The goal of this theory is to show, via rigorous mathematical means, how the performance of EAs depends on their parameter settings and the characteristics of the underlying fitness landscapes. Initially, runtime analysis of EAs was mostly restricted to simplified EAs that do not employ large populations, such as the (1+1) EA. More recently, the theory has been extended to cover complex evolutionary algorithms on realistic problems.

This tutorial gives an introduction to runtime analysis, focusing on methods for runtime analysis of population-based evolutionary algorithms. The tutorial begins with a brief overview of the population-based EAs that are covered by the techniques. We recall the common stochastic selection mechanisms and how to measure the selection pressure they induce. The main part of the tutorial covers in detail widely applicable techniques tailored to the analysis of populations.

To illustrate how these techniques can be applied, we consider several fundamental questions: When are populations necessary for efficient optimization with EAs? What is the appropriate balance between exploration and exploitation and how does this depend on relationships between mutation and selection rates? What determines an EA's tolerance for uncertainty, e.g. in form of noisy or partially available fitness?

Prof. Erik Cambria, College of Computing and Data Science, Nanyang Technological University, Singapore

In recent years, AI research has showcased tremendous potential to impact positively humanity and society. Although AI frequently outperforms humans in tasks related to classification and pattern recognition, it continues to face challenges when dealing with complex tasks such as intuitive decision-making, sense disambiguation, sarcasm detection, and narrative understanding, as these require advanced kinds of reasoning, e.g., commonsense reasoning and causal reasoning, which have not been emulated satisfactorily yet. The Seven Pillars for the future of AI address these shortcomings and pave the way for more efficient, scalable, safe trustworthy AI systems.

Simon James Fong, and Jerome Yen, University of Macau

In the rapidly evolving world of finance, harnessing computational intelligence through advanced machine learning techniques is crucial for gaining a competitive edge. This tutorial will delve into the integration of fuzzy reinforcement learning (FRL) within fintech, offering participants the opportunity to explore leading-edge methodologies for financial analysis and trading.