Actualized intelligence: Case-based agency in practice

Acknowledging the fact that exhaustive search for a plan prior to execution is an intractable problem, planning research has changed course. For the past decade, the study of planning has been recast as the broader problem of interacting with the world, rather than preplanning based on a complete model of the world. This shift in planning proposes agents that work in everyday worlds that must constantly monitor expectations and adjust to contingencies. The Case-based Agency projects at the University of Chicago AI Laboratory demonstrate how reasoning from memory offers a foundation for this approach. The basic insight of case-based reasoning that planning from memory is faster, cheaper, and conducive to learning, has been successfully incorporated into three autonomous agents all working in diverse, but everyday, worlds. This paper shows how the projects implement three techniques for leveraging plans in memory for reuse in everyday worlds. Stabilization is the effort of an agent to impose order on the world so its standard plans are always effective. Learning from failure is the process of recovering from expectation failures, and then using this experience to improve plans in memory. Regularity-seeking perception is an effort to constrain the features in the world that index and monitor plans based on the inherent structure of the world.