This investigation builds new statistical artificial intelligence and reinforcement learning techniques for large-scale manufacturing operations. The primary scientific question it attacks is: How can machine learning be used to automate the decisions that a process engineer makes in monitoring, optimizing, and fine tuning processes? The products of the research are new algorithms for memory-based learning. Innovations include Bayesian locally weighted regression, Fast real-time learning response by means of multiresolution memory- base structures, novel model-selection search algorithms, and experiment design methodologies for memory-based approximators. In parallel with their development, these algorithms are being tested and deployed through several joint projects between the principal investigator's research group at CMU and two large US manufacturing companies. The economic impact of this research will be processes and factories that analyse their own datastreams in real time, develop increasingly improved models of themselves, and exploit those models.
