PUB - Publikationen an der Universität Bielefeld

The present invention uses context information, in order to control an incremental learning process of a system 1 for controlling an automated device even in an unexpected environment. The system 1 is able to correct errors, which are produced by a pre-trained mapping unit 4 that maps an input obtained by primary sensors 2 onto a first control signal 4a for controlling the automated device. The context information obtained by context sensors 3 helps to constrain the learning process of a learning unit 8 to situations, which can be uniquely identified and cannot be covered by the mapping unit 4. A control signal 9a of the combined control signals 4a and 8a of the mapping unit 4 and the learning unit 8, respectively, is computed in a fusion unit 9 by evaluating the confidence values of the two control signals 4a, 8a, and by either choosing one of the two control signals or interpolating between the two control signals. The confidence value evaluation can be done by quantizing the input space according to the incremental learning data, and using the quantized regions for evaluating the performance of both the learned and pre-trained unit within each region.