eunite

research

applications

knowledge transfer

e-archives



eunite 2001A Look Back

Introduction 

 

eunite 2001 - the European Symposium on Intelligent Technologies, Hybrid Systems and their implementation on Smart Adaptive Systems took place from 13-14 December 2001 in Puerto de la Cruz on Tenerife Island in Spain. I was the first symposium organized by the European Network of Excellence EUNITE.


The symposiums' aim was to show new developments in the so called Intelligent Technologies: fuzzy logic, neural networks, evolutionary computation and machine learning. The focus was on hybrid systems that combine two or more of these technologies and/or address issues on adaptivity, preferably with special interest to solve industrial problems. One objective was also to investigate Adaptive Intelligent Systems: systems that coordinate perception, reasoning, and action to pursue multiple goals while functioning autonomously in dynamic environments.
103 papers where submitted to eunite 2001 and 8 authors were invited to present a paper during the symposium. After the papers were reviewed by the chairpersons of the different EUNITE committees, 21 had to be rejected. The remaining 90 papers (82 + 8 invited papers) were presented in parallel sessions.

 

The Programme

In the opening session on 13 December 2001 from 9:00 to 10:30 h, the EUNITE co-ordinator Professor Kauko Leiviska from the University of Oulu in Finland welcomed the 117 participants and introduced them to the symposium. He was followed by the EUNITE project officer Dr. Pekka Karp from the European Commission in Brussels who presented the 6th Framework programme (FP6) (.ppt-file).

The next speaker, Dr. Davide Anguita, chairman of the SAS committee of EUNITE from the Universita degli Studi di Genova in Italy, talked about 'Smart Adaptive Systems: State of the Art and Future Direction of Research'.

The Competition

The opening session was closed with the award of the 1st, 2nd and 3rd winner of the EUNITE competition, which was organised in 2001 by the Technology Transfer Committee. The topic was Electricity Load Forecast using Intelligent Adaptive Technologies. Webpages with full description of the competition can be found on http://neuron-ai.tuke.sk/competition.

 


Organisers and winners of the EUNITE 2001 competition

The chairman of the committee, Dr. Jens Strackeljan from the Technical University Clausthal in Germany, and the organiser of the competition Professor Peter Sincak from the Technical University of Kosice in Slovakia, awarded:

  • the 1st winner Chih-Jen Lin from the Department of Computer Science at the National Taiwan University,
  • the 2nd winner David G. Esp from NGC - National Grid Company in the UK, and
  • the 3rd winner Dr.-Ing. Werner Brockmann from the Medical University in Lubeck, Germany

The winner, Chih-Jen Lin, stated about the competition: "EUNITE 2001 World wide competition provides a good opportunity for researchers and practicians to test their approaches. Being a machine learning researcher, this is a wonderful experience for us to apply theoretical techniques to a real world problem. In addition, we also have to take care every aspect of the data which is very interesting. The organizers carefully prepared the data set and designed the competition. Overall this is a very successful event which I strongly encourage people from different areas to participate it in in the future."

 

David Esp is a Research Engineer involved with smart computing methods in the Modelling Development group at National Grid in the UK. He was encouraged to enter the competition by his clients - the company's Operational Forecasting and Information group, who are responsible (among other things) for the mainstream demand forecasting activity. "We had recently finished the first stage of a project to investigate the applicability and potential benefits of Adaptive Logic Networks (ALNs) within the framework of National Grid's already highly accomplished demand forecasting process". ALNs were identified as the best way forward following a year-long experimentation with a variety of techniques such as neural networks. Much support and advice was provided by their inventor, Prof. William Armstrong, now President of Dendronic Decisions Limited, who provides information and trial downloads from the website. "The competition provided a timely opportunity to show how easily ALNs could be applied to a foreign load forecasting situation. The fact that we managed to apply ALNs to the Eastern Slovakia example in just over two weeks demonstrates just how flexible they are," said David, who thanked his colleague Ian Punch for his help in preparing much of the data for the competition entry. "If only there had been an extra day's time, we could have used bagging, making our results even better, reducing MAPE by about 10%". The conference, apart from offering a break from the English winter, gave David a chance to meet some of the world's leading authorities on a variety of smart computing technologies applied not only to demand forecasting but also to other areas in engineering, finance, health and education. "Many of them could have relevance to National Grid in the future", he said.

"The 2001's EUNITE competition was a very good event, and I enjoyed participation very much", says the 3rd winner Dr.-Ing. Werner Brockmann.. To my opinion, such competitions are very important for the research community as well as for potential users in industry for several reasons. From the research point of view, a competition (on an international level) is a very good opportunity to relate an approach or method and one's own thinking to that of other specialists by tackling the same problem by different approaches. This allows a more dedicate discussion and reveals the properties of an approach more clearly and more dedicately than presenting and discussing only a single successful application of an approach or method. A competition hence gives a more stringent and valuable feedback and allows a broader and more solid rating of different approaches. From the industry point of view, a competition offers a quick, condensed and coherent overview of the current state of the art which is given by the experts of the field. In this way, the company which introduced the problem to be solved reveals the potentials of the field and gets an application-specific ranking of potential solutions for low costs (although the solutions are probably not directly applicable). On one hand it only has to spent some organizational efforts and eventually some prices. On the other hand it gets direct contact to (leading) experts for free. Furthermore, different solutions are worked out at nearly the same time. Hence they are fresh in mind when presented. A competition is thus an excellent way to foster an actual discussion on a methodological level for researches and to push a specific application domain at the same time for industry. Of course, it is important to tackle a real-world problem, instead of an academic (toy) example. Consequently I would like to encourage both sides to use competitions also to foster communication between research and industry.

If you are interested in future competitions, please ask the EUNITE Service Center to be added to the mailing list!

Plenary Session

The purpose of the plenary session was to provide a forum for a general discussion on smart adaptive systems. 5 speakers from different areas were asked to present their work as related to issues of adaptivity. Most of these speakers came from outside of EUNITE and thus provided fresh external perspectives.

In order to focus the discussion, all plenary speakers were asked to address the following three questions:

Anthony Jameson presented his views from the area of user-adaptive systems. He specifically asked the question when and how much adaptivity should be applied. Adaptivity can improve accuracy, but sometimes no adaptivity can be better. Using Bayesian networks, this question can be given a principal answer - something that should be transferable to other machine learning approaches, as well. Finally, Jameson addressed the issue of transparency of models. Usually, a tradeoff is seen between accuracy and transparency. The argument here was that transparency can increase accuracy, while it is in the eye of the beholder what is seen as a transparent solution. The paper he submitted for the proceedings can be downloaded here.

Pete Edwards presented the issue of adaptivity in the context of information agents, providing a perspective from the network of excellence Agentlink II. He noted that adaptivity is not always a defining characteristic of an autonomous information agent, but it plays a large role while being looked at in a great variety of ways. Generally it is defined as responding to changes in preferences or needs, or responding to a changing environment, all with a minimum of user interaction. A new special interest group in Agentlink II is devoted to "agents that learn, adapt and discover". Edwards concluded with saying that "truly" adaptive systems do not exist yet, since most agents often still require a direct user involvement or can exhibit catastrophic break-downs when large change occur.

Mannes Poel presented work on embodied virtual agents (avatars) that interact with the user. He noted that an adaptive system should be able to change or reconfigure its structure during operation in order to better react to the environment. He also pointed out the important role of affective (emotional) component in the context of their work. The important thing is that adaptivity must be embedded in a larger system and must interact with other levels (e.g. BDI). Poel's conclusion also was that truly adaptive systems do not exist yet but that there has been progress. Have a look at the paper 'Embodied Agents in Virtual Environments - The AVEIRO Project'!

Rolf Isermann started his presentation on adaptive control by saying that adaptive systems exist - they can be found in the automatic control community. It must be seen as adjusting behavior in response to changing external influences according to specified performance criteria. Past and present behavior, as well as present performance play a role in adaptation. Isermann presented the theoretical framework of adaptivity in the control domain. Adaptation can be with or without feedback relying on a reference model or model identification. For the latter, excitation is needed to make adaptation possible. A learning system can be defined as adaptation + a goal + decision making. Automatic control systems work quite well, but in exceptional situation they can fail. Thus we need more fault-tolerant systems. Isermann concluded by presenting an example using his own approach - a dynamical neural network (RBFN). If you are interested, please have a look at the paper Modelling and Adaptive Control of Combustion Engines with fast Neural Networks.

Martin Riedmiller presented work of his group in robotic soccer (simulated environment). Adaptation is based on reinforcement learning, through which the robotic agents can learn to adapt to their opposing teams.

The subsequent discussion brought up a number of additional interesting issues:

  • it was noted that much adaptivity might rest within the users rather than in smart adaptive systems
  • it was asked how important control theory is for the field of smart adaptive systems, and why apparently there many problems appear to be solved.
  • in this context it was suggested that learning must be seen on several levels. Control, for instance, is a relative low level.
  • it was noted that adaptivity means learning of mistakes, which is very problematic in the domain of safetly-critical systems. But also in other domains, self-tuning systems might be dangerous.

All in all, it was seen that the different views on smart adaptive system in EUNITE appear to converge, but that much work is still needed (e.g. on taxonomies of concepts) to arrive at a common language when talking about adaptivity.

 

Proceedings and Future Events

The CD-ROM with the proceedings of eunite 2001 is available at the Service Center.

The price for the CD including the booklet with the abstracts of the papers
is EUR 50,-.


If you just want to order the CD, the price is EUR 25,-.

The next symposium, eunite 2002, is planned for 19-21 September 2002. Details can be found here!

news

what is eunite?

how to join?

roadmap

task forces

service center

related links

search eunite with

google



more information

  home