Presentation Team Research themes Projects Conferences Publications

Search themes

There are three main themes:


By learning, we mean here machine learning, namely all the "example-driven" knowledge induction techniques. This ranges from supervised learning in proposition logics or first order logics to unsupervised learning or even adaptative learning.

In the framework of proposition logics, we imagined, implemented, tested and linked to knowledge acquisition models a whole class of learning systems (CHARADE, ENIGME), based on an algebraic formalization using Galois correspondances.

In the framework of first order logics, we reconsidered the notion of structural matching and made effective on a large examples base, introducing a semantic bias related to the a priori structure of the examples. To express those structures, we used the notion of "morion" (REMO). Those ideas are now implemented in the Inductive Logic Programming formalism and extended to account for multi-instances learning (RE...). Finally, specific technics have been developed for sequential data.

Using the notion of generalization spaces, unsupervised learning has been applied to sets of structured objects figured by Sowa graphs (COING system). Granting few restrictions on one to one object matching, some grouping heuristics can be programmed, allowing for an interactive classification of very large set of objects.

Machine learning being mainly based on one reasoning modality, induction, which is now well known, we also worked on setting up a bridge between the old logical and philosophical theories of induction and the more modern and technical approaches of computer science.


Machine discovery uses these different techniques and others, either for datamining or for the rational reconstruction of historical scientific theories or for the discovery of new theories.

The different techniques of supervised or non-supervised learning developed in the group ACASA have been applied or are being applied to datamining (chinese caracters and activity of chemical molecules). The more specific induction technics on sequences have been used to look at recurent patterns in music.

In scientific discovery, we first worked on rational reconstruction of old medical theories. As an example, we studied the origin of the difficulties that preceeded the discovery of the causes of scurvy or leprosy, using the inductive methods developped in the group. We are currently working on physics and electrotechnics. In one case, we want to induce general laws of the physical world using elementary observations made on object behaviour; in the second case we wish to generate useful circuit topology using qualitative reasoning.


Machines creativity often appears too daring to be seriously studied. We did it however on actual cases, both modest and sufficiently complex to be more than a task that could be mecanicaly reproduced with a preprogrammed behaviour. The underlying principle is general : imagination and creativity come from the recombination of past experience stored in memory. Therefore, it is a type of induction since it goes from the specific to the general and, as a consequence, it is a kind of learning as well.

Building intelligent interface agents aims, among other things, at aiding the user. Actions must be suggested to the user according to his/her previous reactions. The SAGE system uses a short term learning called "amnesic learning" because it suggests to the user new tracks inspired by the one he has already explored.

With games, that is, problems dealing with several actors, the one who wins is often the one who anticipated others' reactions. Creative game strategies are automatically generated by a program that modelize its opponents to forsee their reactions and counter them if possible.

A model of improvisation for the bass line of a jazz rythmic trio has been designed and implemented. It uses at the same time a musical memory and some general knowledge on harmony. The principle is based on two mechanisms : one that evocates previously memorized music pieces and one that combines those pieces. To improve the evocation mechanism, that is to set an index in the musical memory, we worked on the discovery of recurent patterns in musical sequences.

Computer Science Laboratory of Paris 6, ACASA team,
8 rue du capitaine Scott, 75015 Paris, Tel: (33) 1-44-27-37-27

-Contact the website manager-