Effector-based goal and operator construction: A model for the design of effector adaptive planners

This thesis addresses two challenges in applying AI planners to complex, practical domains: (1) making AI planners easier to maintain, thereby allowing domain experts to customize and maintain the planner without Al training and (2) allowing AI planners to identify goals from a problem description based on their domain capabilities. These objectives are achieved with a new model for representing and reasoning about the actions a planner can take called Effector-Based Goal and Operator Construction (EBGOC). The salient feature of this model is that the representation is based on effectors, the objects that enable actions, rather than the actions those effectors can produce. An effector-based representation provides a familiar, intuitive interface for domain experts because it encodes knowledge in a form that resembles the objects with which they interact. Additionally, the representation eliminates the fragmentation and redundancy of effector property knowledge found in action-based representations. The result is that the domain experts are able to maintain and customize the planner. Given a problem to solve, the construction engine determines how the effectors can be combined to achieve conditions from the problem description. Operator descriptions are generated to describe these actions and are used to define the set of goals the planner can achieve. This allows the planner to extract its own goals based on the capabilities of the available effectors. The thesis includes an implementation of the EBGOC model for the domain of machining process planning. The implementation, called MEDIATOR, identifies machining features and operations given CAD models of the part and stock. Evaluations show that MEDIATOR is easy to customize and maintain, able to support a computerized setup planner and generates machining options with quality that meets, and in some cases, exceeds those identified by domain experts. These results indicate that the EBGOC model can be successfully applied to complex, practical domains.