Abstract
This thesis deals with constraint network acquisition, which consists in automatically acquiring a constraint network formulation of a problem from a subset of its solutions and non-solutions. Our work is mainly focussed on the constraint network acquisition platform CONACQ. In the standard version of \conacq, the choice of the subset of solutions and non-solutions to use for learning was assumed to be made before and independently of the acquisition process. In the first part of this thesis, we present an interactive version for CONACQ in which the acquisition system actively assists in the selection of the set of examples used to acquire the constraint network through the use of learner-generated queries. We show that the number of examples required to acquire a constraint network is significantly reduced if queries are selected carefully. In order to decrease the number of interactions, we provide a theoretical and empirical evaluation of query generation strategies for interactive constraint acquisition, with very positive experimental results. In the second part of this thesis, we are interested in a pratical use of constraint network acquisition for robotics. Thus, we propose a Constraint Programming-based framework for modelling, planning and supervising sensorimotor behaviors. Our approach uses the constraint network acquisition platform CONACQ to automatically encode by Machine Learning each elementary action of a robot as a CSP. The acquired CSPs are then automatically combined by planning to constitute sensorimotor behaviors using a CSP-based planner inspired by CSP-Plan. Moreover, our architecture exploits the propagation properties of the acquired CSPs to supervise the execution of a given sensorimotor behavior. Some experimental results are presented to validate our approach.