The paper "Multi-Plan Retrieval and Adaptation in an Experience-based Agent" by Ram and Francis presents a theory of "experience-based agency" and uses an implementation of this theory as a tool to investigate the use of multiple plans in case-based planning. Salient points of the theory include the focus on real-world complexity and a holistic approach to handling this complexity. The authors' proposed architecture includes all aspects of the planning task in a complex system informed by topics in software agent theory. That is, the architecture is modeled as an intelligent agent with dynamic internal modules to facilitate memory, reasoning, and agent control.
To understand the architecture, we need to first explain least-commitment planners. Least-commitment planners solve the problems by successive refinement of a partial plan. Because it adapts only one plan, it can resort to significant amounts of from-scratch planning. To make the most effective use of the planner's past experience, we need the ability to recognize when a partial plan needs to be extended, the ability to select plans that address the efficiency, and then extract and merge the relevant parts of the retrieved plan into the original plan.
MPA provides this ability by extending least-commitment planning with three components: a goal deriver, a plan clipper, and a plan splicer. These mechanisms provide the ability to decide what experiences to combine and when to combine them. NICOLE-MPA was implemented to show that plan adaptation should be driven by retrieval and the resulting MPA implementation indicated significant speedup over previous least-commitment planning algorithms.
Francis and Ram compare NICOLE with several other systems... frequently determining that the system is neither as asynchronous nor dynamic as NICOLE. For example, macro-operators are precomputed at storage time, not like Clippings, and systems like Prodigy/Analogy (ch. 8) store totally ordered plans. Francis and Ram believe this causes difficulty attempting to interleave steps, while the least-commitment strategy of the MPA algorithm avoids such problems.