| Chemical process design is a complex, open-ended activity that requires the design agent to possess abilities currently missing in most artificial design systems. The ability to integrate and apply the diverse knowledge sources required for process design and the ability to learn from experience are two important capacities. Soar is an integrated software architecture for knowledge-based problem solving, learning and interaction with external environments.; This thesis reports on two systems, both developed within the Soar architecture, that display the above abilities. The first, CPD-Soar, designs distillation sequences, while the second, Interval-Soar, performs simple arithmetic tasks. In describing the structure and behaviour of both systems, it is depicted how design and design-related activities can be cast within the Soar framework. The systems collectively provide evidence for an hypothesis about learning; namely, that the knowledge learned by an agent while performing a task is strongly dependent upon the model the agent brings to the problem-solving experience. Specifically, it is shown that the richer the models an agent has of its evaluation functions, the more general the knowledge it learns.; The thesis also describes CPD2-Soar an enhanced version of CPD-Soar that also incorporates the functionality of Interval-Soar. CPD2-Soar depicts how the selection of an evaluation function can be posed as a problem-solving activity within a process designer. It also depicts how a process designer can learn to improve its evaluation-function choosing ability by abstracting from its problem-solving experiences. The structure and expected behaviour of CPD2-Soar are postulated.; The thesis also introduces the notion of marginal price, the change in price of a chemical separations task as a result of performing it in the absence of non-key components. The marginal price can be computed using any metric that reflects process economics; vapour flowrate and total annualised cost are typical examples. The utility of the metric in quantifying and explaining a number of heuristics used for distillation column sequencing is described as is its performance as an evaluation function for controlling the search in distillation-sequence design problems. |
