Machine Learning

Being able to build computer systems that can learn in some sense is one of the very central problems in artificial intelligence. Inspired by certain theoretical advances in the field such as Valiant's PAC model and Rissanen's MDL principle, the project generally aims to apply to machine learning the approach of theoretical computer science and algorithmics. This international trend is called computational learning theory. Our theoretical work is supported by experimenting with the algorithms.

Results on decision tree learning include analyses of the requirements for efficient multisplitting on numerical attributes and comparison of the most commonly used attribute evaluation functions with respect to these requirement. A practical method that results in optimal splitting has also been developed. Testing environments for learning algorithms have been developed.

Some promising results have been obtained about MDL based learning algorithms in particular for the case where the instances are strings of arbitrary length. The method has been applied to a clustering problem of biological sequences.

Research has also been done on on-line learning algorithms that make no assumptions about the distribution of noise in the data. This is known as agnostic on-line learning. Emphasis has been on learning simple statistical models, such as generalized linear models, with algorithms that learn fast even if there is a large number of irrelevant variables present.

The current research themes of the group are:

• development of MDL-based and other learning algorithms with a particular emphasis on potential applications in computational biology,
• development of decision tree learning algorithms that have a sound theoretical basis and perform well in practice,
• generalizing the agnostic on-line learning methods further and deepening the understanding of their relationship with more classical statistical approaches,
• applications of symbolic learning approaches to problems of mobile robotics, and
• (lazy) learning for planning in continuous domains, with applications to an industrial case-based bioprocess planner (joint work with VTT and Finnish breweries).

The group has good international reputation and is together with nine other European groups a member of the ESPRIT Working Group NeuroCOLT II and a site of European Machine Learning Network. The group also works in close co-operation with Prof. Heikki Mannila's data mining group. The members of the group are Prof. Esko Ukkonen (group leader), Doc. Jyrki Kivinen, Dr. Tapio Elomaa, M.Sc. Tibor Hegedüs, M.Sc. Markus Huttunen, M.Sc. Juho Rousu (VTT Biotechnology and Food Research) and M.Sc. Jaak Vilo. The group gets funding from the Academy of Finland and from the European ESPRIT Programme.

Publications: [189, 195-197, 203-206, 208-215, 218, 220-222, 259-262].

Home Page: http://www.cs.helsinki.fi/research/pmdm/ml/