|||Martin Spott, Ralf Schoknecht, Martin Riedmiller, Approaches for the integration of a priori knowledge into an autonomously learning control architecture, Proc. of EUFIT 99, Sep 1999.
The control architecture Fynesse learns control strategies only on basis of success and failure of former control interactions. Thus, learning can be carried out directly at the real process. For reasons of safety and design costs, the number of interactions with the real process must be kept as small as possible. Therefore Fynesse allows the integration of a priori knowledge about the control strategy. In this paper, different techniques for the integration are proposed. In experiments it is shown that the integration of prior control knowledge reduces the number of interactions with the process dramatically. Furthermore the learning procedure is stabilized, which is indispensable for the application of Fynesse as an adaptive system.