Building a chemical process design system within soar—2. Learning issues

Abstract We explore the potential to include automatic learning in a design agent by implementing a simple distillation sequencing system, CPD-Soar, within Soar. Soar is an integrated software architecture with a build-in set of mechanisms for exhibiting intelligent behavior, including problem-solving, learning and interaction with the environment. Soar has a number of scientific uses: computer scientists build artificially intelligent agents with Soar as a foundation, and cognitive psychologists use Soar to model human cognition. CPD-Soar illustrates how design-related tasks can be cast within Soar's framework, hence demonstrating the functioning and potential of its problem- solving and learning mechanisms. This simple example system, which involves computations with real numbers, automatically learns things which are too specific, leading to the hypothesis that the generalization an agent infers from specific examples is strongly dependent upon the model the agent brings to the learning process. We introduced the Soar architecture as a vehicle for developing design systems with capabilities seen to be important for chemical process domains but missing in most existing design systems. We reported upon CPD-Soar, a system developed within the Soar framework, and described in depth its tasks, problem-space structure, operation and performance. The construction of CPD-Soar was a valuable exercise for two main reasons: one, it provided evidence that the mechanisms present in Soar can provide design systems with useful abilities, and two, the act of creating the system was useful in distinguishing between those aspects of the task domain that are well understood from those that are not. Selecting among competent evaluation functions and learning within numerically-intensive domains were two areas identified as not being well understood. Our observation of CPD-Soar's learning behavior lead us to postulate an hypothesis about learning: namely, that the richer the model an agent has of its evaluation functions, the more general its learning will be.