ICAISC 2022

With the recently publicized USA’s ambitions to enable human exploration of the Moon and Mars in the next decade, combined with the NASA’s long-standing open data policy, we can observe an increase in opportunities for performing interesting and impactful machine learning research on the rich space weather data, which includes the Sun’s magnetic field data, and other measurements coming from multiple spacecrafts or ground-based observatories. These research opportunities can be further enriched with another exciting source of big data, which comes from numerous heliophysics-related scientific simulations.
We hope to increase awareness about these promising machine learning research opportunities through this seminar. Firstly, we will briefly discuss the impacts of transient solar events on the Earth and space travel to clarify the importance of accurate space weather predictions. While these topics have always been central to heliophysicists’ research interests, they are nowadays intensively investigated by the operational community, involving major federal institutions, such as National Oceanic and Atmospheric Administration (NOAA), the United States Space Force, NASA, and many dynamically growing industries related to communication satellites, air travels, power grids, precision agriculture, oil drilling, and (recently growing) space tourism. Secondly, we will discuss some aspects of freely available big data coming from the Sun-pointing satellites (NASA, NOAA, ESA), multi-institutional networks of ground observatories (e.g., GONG H-alpha Network, hosted by National Solar Observatory), as well as computationally demanding high-resolution solar activity simulations conducted by the heliophysicists-led academic units, governmental and private research labs, as well as small businesses. Finally, we will present some examples of the space weather-related research performed in the interdisciplinary Data Mining Lab at Georgia State University (https://dmlab.cs.gsu.edu/), where we try to serve, through collaborative initiatives, both research- and operations-focused space weather communities with big data-driven machine learning research and applications.

Multi-objective optimization problems (MOPs) are ones that require simultaneous optimization of multiple conflicting objectives to attain the state of Pareto-optimality, where improving solutions in terms of one objective leads to deterioration in terms of one or more of the other objectives. MOPs galore in diverse real-life situations, and several algorithms have been developed for solving them. Multi-modal MOPs (MMMOPs) are those problems where a many-to-one mapping exists from solution space to objective space. As a result, multiple subsets of the Pareto-optimal set could independently generate the same Pareto-Front. The discovery of such equivalent solutions across the different subsets is essential during decision-making to facilitate the analysis of their non-numeric, domain-specific attributes.
In this talk, we will first provide a brief introduction to MOPs and an algorithm for solving them, followed by an application to the real-life problem of drug design. This will be followed by a discussion on the basic concept of multi-modality in MOPs. We then identify a problem of the existing approaches for solving MMMOPs, which is referred to as the crowding illusion problem. A method for solving MMMOPs with a graph Laplacian-based Optimization using Reference vector assisted Decomposition (LORD) will thereafter be discussed. The talk will conclude with the mention of an application of MMMOPs to the problem of building energy optimization.

New approaches to extract useful information from EEG are the key to use this technique for diagnosis, biofeedback and brain-computer interfaces. Techniques based on event related potentials, motor imagery and steady state visually evoked potentials (SSVEP) are still dominating the field, but most interesting developments are in investigation of neurodynamics, observing information flow through the brain connectomes. Many new mathematical techniques have been proposed, but development of methods useful in clinical practice is still a great challenge. I will review most interesting new approaches in this areas and summarize our own attempts in analysis of frequency-based fingerprints, new spatial filters, recurrence-based nonlinear features.

[1] Rykaczewski, K, Nikadon, J, Duch, W, Piotrowski, T. (2021). SupFunSim: spatial filtering toolbox for EEG. Neuroinformatics 19, 107–125
[2] M.K. Komorowski, K. Rykaczewski, T. Piotrowski, K. Jurewicz, J. Wojciechowski, A. Keitel, J. Dreszer, W. Duch (2021) ToFFi - Toolbox for Frequency-based Fingerprinting of Brain Signals (submitted to Neurocomputing).

In a smart grid context, a demand response strategy of electric vehicle charging is modelled by a stochastic game, where a big data analytic framework is proposed for controlling the electric vehicle charging behaviours. We will also look at Plug-In Electric Vehicles (PEVs) Charging: Feeder Overload Control problem. Moreover, a two-stage stochastic game theoretical model is proposed for energy trading problem in a multi-energy microgrid system. Concerning the privacy, a research branch of reinforcement learning (RL) that dominates distributed learning for years will be presented by making the first attempt to apply RL-based algorithms in the energy trading game among smart microgrids where no information concerning the distribution of payoffs is a priori available and the strategy chosen by each microgrid is private to opponents, even trading partners. To solve this challenge, a new energy trading framework based on the repeated game that enables each microgrid to individually and randomly choose a strategy with probability to trade the energy in an independent market so as to maximize his/her average revenue.

Decision making – which is meant as a selection of an option, or a set of options, or a course of action, or courses of action – is presumably the most common and frequent human activity. More and more often decision making proceeds in complex settings in the sense that there are more and more stakeholders, much uncertain, imprecise and missing information, all characteristics and specifications change over time, multiple criteria, goals or aspects are to be accounted for. Moreover, an explict human centricity is pronounced in the sense of a crucial role of the human being in the decision process. These complications clearly suggest that the humans involved in the decision making process should be aided or supported by some computer based tools and techniques.
In most cases a more effective and efficient setting is to assume the decision making to proceed, first, in a setting in which there is a domain expert, the so called judge, who has a deep expertise in his domain, e.g., city transportation, but not in the solution tools, e.g. optimization, and he/she commissions an additional expert, the so-called advisor, who has a deep expertise in the solution methods for some class of formally stated problems but not necessarily in the domain. For such a setting we present the so-called judge-advisor type approaches, and consider in particular issues related to advice giving, taking, and rejecting. We show that AI, notably via machine learning but also via analyses of argumentation or explanation, can help improve this judge-advisor scheme.
Then, we propose as a convenient, „democratic” solution the decision support systems (DSSs) which can be used by less qualified users, and we show that data driven approaches open here new vistas. In this contaxt of a wide use of algorithms for automating many tasks, we consider the so-called algorithm aversion, notably its aspects related to a need for comprehensible and trustworthy results of such approaches, as well as when, for which problems, etc. these new autonomous AI-enabled decision support systems can be used.
Finally, we briefly summarize new qualities that can be obtained by involving the concepts, and tools and techniques of AI-assisted decision making and AI-enabled decision support. Some examples on socioeconomic planning are mentioned.

The talk argues and demonstrates that the third generation of artificial neural networks, the spiking neural networks (SNN), can be used to design brain-inspired architectures that are not only capable of deep learning of temporal or spatio-temporal data, but also enabling the extraction of deep knowledge representation from the learned data. Similarly to how the brain learns time-space data, these SNN models do not need to be restricted in number of layers, neurons in each layer, etc. as it is the case with the traditional deep neural network architectures. When the SNN model is designed to follow a brain template, knowledge transfer between humans and machines in both directions becomes possible through the creation of brain-inspired BCI. The presented approach is illustrated on an exemplar SNN architecture NeuCube (free software and open source available from www.kedri.aut.ac.nz/neucube) and case studies of brain and environmental data modelling and knowledge representation using incremental and transfer learning algorithms These include predictive modelling of EEG and fMRI data measuring cognitive processes and response to treatment, AD prediction, BCI, human-human and human-VR communication, hyper-scanning and other. More details can be found in the recent book: Time-Space, Spiking Neural Networks and Brain-Inspired Artificial Intelligence, Springer,2019, https://www.springer.com/gp/book/9783662577134.

Image classifiers based on convolutional neural networks are defined, and the rate of convergence of the misclassification risk of the estimates towards the optimal misclassification risk is analyzed. Under suitable assumptions on the smoothness and structure of a posteriori probability, the rate of convergence is shown which is independent of the dimension of the image. This proves that in image classification, it is possible to circumvent the curse of dimensionality by convolutional neural networks. Furthermore, the obtained result gives an indication why convolutional neural networks are able to outperform the standard feedforward neural networks in image classification. Our classifiers are compared with various other classification methods using simulated data. Furthermore, the performance of our estimates is also tested on real images.

In this lecture first we will describe fundamental and current trends in Artificial Intelligence (AI), and Data Science. In this regard we will demonstrate applications of different machine learning algorithms including Deep Learning and Graph Neural Network in real life application like Intelligence Car and healthcare. We will discuss the important of explainable and trusted AI. Finally we will also discuss issues and challenges related to Big Data.

Ambient Assisted Living (AAL) is evolving rapidly and has undoubtedly become a topic of great relevance given the aging of the world population. Advances in networking, sensors, and embedded devices have made it possible to monitor and provide assistance to people in their homes. Such an enormous growth of AAL-related AI applications has prompted recommendations that AI techniques should be "human-centered". There is no clear definition of what is meant by Human Centered Artificial Intelligence. Unfortunately, while we can define several components that compose an Human-Centered AI agent, we probably face a quite long road in order to workaround a noticeable series of formal problems in order to obtain such a standardized definition. Europe's diverse AI community are bringing together individuals and organizations that are interested in contributing to, or benefitting from, today's AI capabilities and that are interested on a formal definition and widespread adoption of standards for Human-Centered AI systems. An entire AI ecosystem is nowaday growing moving forward the frontiers of our knowledge and application capabilities on this field.

Granular Computing is about representing knowledge by means of information granules, constructing information granules, their processing, and realizing communication and interpretation carried out in the framework of information granules.

Information granules are abstract constructs that bring together individual entities because of their closeness, similarity, or resemblance. The level of abstraction makes a description of the problem manageable and problem solving strategies feasible and efficient.

We offer a rationale behind emergence of information granules, offer examples and present a variety of frameworks (sets, intervals, fuzzy sets, probabilities, rough sets, random sets, intuitionistic sets…) using which they are formally represented.

Main motivating factors are advocated. General ways of designing and evaluating information granules are discussed. A role of a variety of clustering techniques treated as a prerequisite for the formation of information granules is demonstrated. The evaluation of the quality of information granules is case ion the granulation-degranulation scheme.

We deliver a comprehensive approach to the development of information granules by means of the principle of justifiable granularity; here various construction scenarios are discussed. In the sequel, we look at the generative and discriminative aspects of information granules supporting their further usage in the formation of granular models. A symbolic manifestation of information granules is put forward and analyzed from the perspective of semantically sound descriptors of data and relationships among data. The principle provides a way to build an information granule such that it is legitimate from the perspective of coverage (experimental legitimacy of the granule) and its semantics (meaning). Along with the generic construct, discussed are various augmentations of the principle. We carefully look at the generative and discriminative aspects of information granules supporting their further usage in the formation of granular artifacts. The considerations are carried out following a general knowledge representation scheme:

data -› numeric prototypes -› information granules -› symbols

Furthermore, a symbolic characterization of information granules is put forward and analyzed from the perspective of semantically sound descriptors of data. Their linguistic summarization is offered as well. The diversity of information granules is also captured by more advanced constructs of information granules of higher type and higher order.

Some selected topics of data analytics in which information granularity is visible such as (i) imputation, (ii) time series prediction, (ii) data stream analysis, (iii) imputation, (iv) association analysis and associative memories, and (v) transfer learning are formulated and discussed.

The tutorial is made self-contained; all required prerequisite material will be made an initial part of the presentation.

Undoubtedly, it can be said that nowadays, evolutionary algorithms (EA) or so-called evolutionary computational techniques (ECT) represent the most popular metaheuristics that allow solving complex optimization problems. Over time, two large groups of algorithms have formed, distinguished by different philosophies of function (mainly due to different inspirations in nature). Thus, in the literature and the professional community, one can find these two essential groups labeled as "classical EA" and so-called "swarm" algorithms. Both groups of algorithms, together with recently published findings and results, carry specific challenges and also criticisms.
This talk aims to focus specifically on the recent challenges in the ECT field, algorithm development trends, and analysis that could eliminate many of the problems and resolve the open challenges. In recent years, incremental algorithm performance and efficiency advances have been made primarily on benchmark sets, but benchmarking is not always done correctly. Thus, the recent trend of so-called good practices in benchmarking will be discussed. Furthermore, possible approaches and SW tools (frameworks) for algorithm autoconfiguration, i.e., the appropriate setting of algorithm hyperparameters for given instances of the optimization problem. Another important topic that will be mentioned is also the possibility of linking the development of ECT and explainable AI, in particular, how the notion of explainability can be understood in relation to ECT.
The lecture is designed to be self-explanatory, without requiring knowledge of ECT theory.