Putting the Problem Solver Back in the Driver's Seat: Contextual Control of the AMTS

The atms is a powerful tool for automated problem solvers and has been used to support several model-based reasoning tasks such as prediction and diagnosis. It provides an eecient mechanism for maintaining consistent sets of beliefs and recording the assumptions underlying them. This enables the problem solver to switch rapidly between contexts and compare them. Such capabilities are central to diagnostic systems, and are also valuable to design and planning systems. Applications to larger problems have been hampered, however, by the problem solver's inability to maintain control over the atms. We present a new approach, implemented in a system called coco, which allows the problem solver to maintain tight control over the contexts explored by the atms. Coco provides means for expressing local and global control over both normal and nogood consumers. Local control is achieved by attaching guards to individual consumers. These guards express control, rather than logical, knowledge and consist of sets of environments. Global control is achieved by specifying a set of interesting environments. Consumers are red only when its antecedents are true in some interesting environment. We also successfully apply the same technique to limit label propagation in the atms. This ensures that the atms respects the problem solver's wishes and only makes derivations in interesting contexts. We demonstrate the both the dramatic increases in eeciency which are made possible by these techniques, as well as their tremendous expressive power, in four examples.