The Covid-19 crisis pushed many of us to change our working styles. As universities all over Germany changed to an online version, video streaming became a daily necessecity for students to follow the lectures and for professor to be able to evaluate the students' presentations. We hear very often that the stay-home orders amend Covid-19 crisis had an immediate positive impact on the air-pollution and the environment in general but the question is if our new adapted life-style keeps this positive impacts also on the longer runs. Therefore, it is very interesting to answer questions like: "What are the CO2 costs of video streaming?". Professor Wolfgang Konen takes a look at this very important question in this article (unfortunately only in German).
I am very happy to announce that on Wednesday, April 8th 2020, the PhD-colloquium of Samineh Bagheri, which I had the honor to supervise at TH Köln, successfully took place. Her thesis “Self-Adjusting Surrogate-Assisted Optimization Techniques for Expensive Constrained Black Box Problems” deals with state-of-the art optimization algorithms supported by RBF surrogate models. In normal times we would have been all traveling to Leiden University and would have made the ceremony in the prestigious Leiden senate hall. But times are not normal. Due to Covid-19 it was impossible to proceed as planned. But Leiden University is very innovative, and they decided to do their first-time ever fully-online Weiterlesen »
We are happy to inform you that a new version of the SACOBRA package is available on CRAN. SACOBRA is a self-adjusting surrogate assisted optimization framework designed for expensive black-box constrained problems.
This version of SACOBRA can handle black-box equality constraints as described in this paper. Also the online model selection algorithm described in an award-winning paper is available to use in our recent release of SACOBRA. Additionally, the implementation of all 24 G-problems (the most commonly used test-suite for black-box constrained optimization problems in optimization community) is directly available in our package. If you have used SACOBRA 1.1 you might have experienced repeated warning messages coming from nloptr package, this issue is also resolved for the moment.
You can easily install SACOBRA by running the following line in your R command line:
install.packages("SACOBRA",dependencies=T)
If you have any interesting or challenging constrained optimization problems that you want to solve efficiently, apply SACOBRA and let us know about your experience.
The GroupLearn proposal of Prof. Konen is one of the three proposals from TH Köln that wins a Stifterverband Fellowship in 2020. This year only 10 of the 87 submitted Senior Fellowship proposals were selected for funding by Stifterverband. These fellowships were awarded with 50.000 Euro funding.
Stifterverband NRW awards Fellowships for Innovations in Digital University Teaching for purposes like redesign of university teaching modules using digital technologies and the development of digitally supported teaching and examination formats.
GroupLearn - Group-based learning in the digital learning world MathWeb
In order to promote group work among students, Prof. Wolfgang Konen wants to create with "GroupLearn for MathWeb" a space where students can support and exchange ideas. "GroupLearn", is a digital learning world that is intended to provide a fun and impactful environment to practice mathematical problems. Two main functionalities of GroupLearn can be listed as follows:
- The Exercise Compiler is designed to provide an interactive, user-friendly environment in which students design new exercises and their solutions for MathWeb that other students can solve. (The picture above shows a possible layout of the exercise editor.)
- The Discussion Forum is intended to provide a place for students and learning coaches to exchange solutions, questions, and problems.
Electrocardiogram with Anomalies
Markus Thill, who is PhD student in our CIOP group, is working on developing new anomaly detection algorithms since 2017 under supervision of Prof. Wolfgang Konen. The field of anomaly detection, which is today often tackled with machine learning algorithms, became a hot topic in the last years. Thill's recent paper titled "Anomaly detection in electrocardiogram readings with stacked LSTM networks" won the best paper award at the ITAT conference. LSTM (Long Short-Term Memory) networks are a special form of recurrent neural networks and thus belong to the group of deep learning algorithms.
ITAT (Information technologies -- Applications and Theory) is an annual European conference being held in Slovakia. This year ITAT took place from 21 till 23 of September and hosted a number of interesting talks focusing on time series analysis topics including anomaly detection, forecasting etc. " Word guessing game with a social robotic head " was one of the three invited talks in this conference.
Thill uses a deep learning approach to tackle the challenging task of anomaly detection. This method, called LSTM-AD, is described in detail in this paper and shows very promising results compared to the state-of-the-art (see figure above). A distinguishing feature of LSTM-AD is that it learns unsupervised what an anomaly is, just by analysing sufficient data with predominantly normal behaviour. Thill's method is applied to electrocardiograms (heart beat recordings) and can detect anomalies in these electrocardiograms with great reliability.
The Many Criteria Optimization and Decision Analysis (MACODA) workshop, organized by a team of successful researcher and scientists in the field of multi and many obejctive optimization was held from Monday 16 September through Friday 20 September 2019 in Leiden.
The motivated participants of this workshop including early stage researchers, post docs, professors and several researchers working in industry all had a shared goal and that was to address this question: "How should we use computational support to deal with problems with many objectives?" This question is crucially important, as the real-world optimization problems can often be formulated as multi or many objective optimization tasks.
The first day of the workshop belonged to several interesting talks given by pioneers in the field of many and multi-objective optimization, Prof. Hisao Ishibuchi and Prof. Miettinnen. Also, the early stage researchers found the opportunity to share their recent works. From the very first day participants found time and opportunity to discuss different topics and challenges in the field. The atmosphere of sharing and learning from each other got even better as we started with the open space workshop model on the second day. In this form of workshops everybody is invited to come up with topics that have a high importance to discuss with the community. Everyone who comes up with a topic can write it on a card and attach it to the wall in a specific cell showing a specific time slot and room. This is an official invitation to the rest of the participants to join the discussion if it is of their interest. Every discussion was documented.
In the course of the workshop a lot of existing challenges in the field along with possible cures were discussed. Some of the groups ended up finalizing their work with a plan for future collaborations in form of a paper or a project.
Wolfgang Konen presented his work on General Board Game Playing at CoG'2019, the Conference on Games, which took place between 20-23 of August in London, UK. You can find the poster here and the accompanying paper here on arXiv. The General Board Game (GBG) Playing Framework is about computer agents that learn game strategies just from self-play and work across a large variety of board games. It is also an educational project that makes it easier for computer science students to start their projects, Bachelor or Master theses in the area of game learning. The GBG poster presentation led to a series of interesting discussions and feedbacks.
Conference CoG (formerly known as Conference on Artificial Intelligence in Games) started with a fascinating keynote given by David Silver from Google's DeepMind. David shared very interesting insights about their famous AlphaGo and AlphaZero learning agents and their new works AlphaStar (mastering the real-time-strategy game StarCraft II) and AlphaFold (cutting-edge technology for 3D protein folding). Many other interesting talks, posters and tutorials (among them one given by Boris Naujoks, TH Köln, together with Vanessa Volz on Game Benchmarks for EAs) made the conference a very stimulating event. Apart from interesting keynotes, talks and discussions, a "Games Night" event was organized which spiced up the whole program.
This year, beautiful Prague hosted GECCO 2019, from 13th to 17th of July. GECCO -- standing for The Genetic and Evolutionary Computation Conference-- is an outstanding conference which brings many active scientists and PhD students together annually. We have presented a recent piece of our work in form of a poster. The work titled: "Online Whitening for High Conditioning Optimization Problems" is a preliminary work toward addressing the challenges of handling high conditioning functions with surrogates. GECCO provided a great opportunity to recieve interesting feedback and ideas from the optimization community.
From the 501 papers received, 173 contribuations were accepted as full papers and 168 contributions as poster by GECCO 2019 committee. High quality paper and poster sessions both enabled valubale chances to exchange opinion and evolve your ideas. Apart from these sessions, GECCO hosted plenty of useful tutorials, workshops, keynotes and social events.
"Challenges for Learning in Complex Environments" was the name of an outstanding keynote talk given by Raia Hadsell, the Robotics Lead of Google DeepMind. She emphasized on the importance of multiobjective learners and the continual learning concept.
We are pleased to inform you that the recent work "General Board Game Playing for Education and Research in Generic AI Game Learning" of Prof. Wolfgang Konen has been accepted for publication and it will be presented as a Poster at IEEE CONFERENCE ON GAMES (CoG) 2019. CoG 2019, held in London, UK, will be hosting many peer-reviewed paper and poster presentations as well as academical and industrial invited talks in various game technology related fields. Wolfgang Konen introduces a general board game playing and learning framework, in his recent paper. You can find more information about the developed educational tool by Wolfgang Konen in one of our former blog-posts.
Recommender System is the name given to any software designed to recommend you objects to purchase, to click on or to watch, which might be attractive to you with a high probability. Development of such systems became a hot topic for software developers and machine learning engineers in the last years. A well-performing recommender system can make significant contributions to various online platforms.
Nowadays, many of us interact with recommender systems on a daily basis. If you have an account in Amazon or Netflix or many similar online services, you may have received suggestions that often fit your interest. This can vary from a movie suggested by Netflix or very various products suggested by Amazon. The targeted advertisements on Facebook or the targeted pop-up advertisements in your browser are also other examples of applications of these systems. Almost no two person receives the exact same advertisements or suggestions on an online platform. All these personalized suggestions and advertistments use recommender systems to find the best match for your interests according to your former ratings, purchases or clicks.
Content-based recommender systems often make use of the explicit features of the users and the items to provide recommendations. In contrary, the Collaborative Filtering recommender systems do not have access to any sort of explicit features and tend to learn the important features of users and items implicitly. The real-world recommender systems are often a hybrid of both approaches.
You can try our movie recommender system demo, by simply giving your own ratings and see what the system will suggest to you. This movie recommender system is based on the implementation of Matrix Factorization algorithm in Python. Matrix Factorization is a collaborative filtering approach which tends to learn implicit features for users and items by factorizing the exiting rating matrix. We expalin this algorithm in more details, in this Google Colaboratory Notebook . We describe how the recommender system problem turns into an optimization problem and how to handle such optimization tasks. Also a step-by-step implementation of this algorithm in Python is provided.