Up to date publication list is available at my GoogleScholar page
Subject areas: Computer Science, Learning Systems
Application areas: Intelligent Vehicles
Projects: ReDi2Service, VPMS, InnoMerge, FuelFEET, EIS-IGS Smart Grids, Science without Borders, In4Uptime, ARISE, IMedA, SeMI, HEALTH, BIDAF, EVE
- Artificial Intelligence
- Machine Learning and Data Mining, especially for Streaming Big Data
- Knowledge Representation
- Rational Agents
- Planning and Non-standard Logics
- Data-driven fault detection methods
I have first became interested in Artificial Intelligence and especially Machine Learning during my Master studies, because it appeared to be the ultimate solution to all the computational problems. Since then I have spent most of my career trying to understand the actual limitations of those methods. During my PhD studies I have primarily focused on rich knowledge representations and connection between learning and non-standard logics. More recently I am focusing on self-organising and unsupervised approaches. My main interests lie in the data-driven modelling and knowledge representation.
Coming to IS-lab at Halmstad University, I have joined a research track that focuses on knowledge discovery in large and distributed data streams, including unsupervised modelling. An example is the case of heavy-duty vehicle diagnostics, where ground truth is sparse or even unavailable. The goal is to develop new, or adapt existing, Machine Learning and Data Mining techniques to make them suitable for handling massive streams of data in a distributed and self-organising manner. In the spirit of "aware systems" that form the foundation of CAISR environment, my research concerns systems that can be human-, situation- and self-aware.
One way to achieve this goal is through discovery of interesting patterns and relations, where the "interestingness" can be treated as a metric and quantitatively measured. In this respect we are evaluating of both the data and the extracted knowledge. In many applications it is not feasible to store all the data, and therefore a preliminary decision needs to be made as to what are the most useful subsets to use in further analysis. We aim for interestingness metrics that are suitable for evaluating partial results in distributed environments. An important feature, however, is that they should be adaptable to different tasks and domains, as well as work for both supervised and unsupervised learning.
Specifically on the topic of self-organisation and self-awareness, we have had successes in creating solutions that work well for specific application domains. The next step I am interested in is to obtain deeper understanding of fundamental concepts, allowing us to build a general theory on top of those successful application examples.
I am also interested in guiding learning process using (both structured and semi-structured) expert and historic knowledge. In particular, this can be done before the learning starts, but also later, as a way to evaluate results and have the user guide the process in an interactive way. I am working towards designing knowledge representation models that allow for efficient learning, while being flexible enough to capture different aspects of the data simultaneously.
It is evident from the exponential development in e.g. sensors, embedded computers, computing power and memory capacity: all those artifacts are becoming less and less self-centered and their value depends on their ability to fit their environment. We today see the result of this in, e.g., the automotive industry where modern cars are more complex than what the average repair-person can handle (and the "competence" gap is growing). The industry is therefore investing in automatic fault detection and diagnostics, where devices "know" more and more about themselves and their normal operation, providing aids for the maintenance and repair personnel. The same development is occurring in the health care sector, i.e. the technological equipment and the possibilities are more complex than the average care-giver can manage. It is of vital important to have systems that are "aware" and easy to manage.
Overall, I am interested in Artificial Intelligence because I want to make computers do things which are "easy", in addition to those that are difficult: to not only play chess, also soccer. I believe that the way to achieve this are interdisciplinary approaches. At the moment a lot of incredible AI solutions are being created, but there is not enough effort to combine them into a truly intelligent, complete systems. I am convinced that all truly reliable systems must be able to learn from experience, and thus I consider Machine Learning techniques to be necessary. However, they need to move away from some of the very strong underlying assumptions that work in the lab but are not possible to meet in real world. Finally, I am convinced that focus needs to be put on resource-bounded agents and systems. Idealised models such as omniscience are useful to understand basic principles, but we are now ready to move towards more realistic settings.