Enabling mechanisms for AI planning knowledge sharing, merging, and reuse

Abstract

With the availability of cheap, powerful computers and the proliferation of the internet to connect them, recent trends in computing are focusing on peer-to-peer interoperation. From an intelligent systems perspective, cooperation and knowledge sharing between peer systems can enable each peer to overcome knowledge deficiencies and computational ability limitations. This dissertation investigates issues in enabling such peer-to-peer interoperation between domain-independent planning/problem-solving systems. The dissertation is divided into three primary topics in planning/problem-solving: (i) multi-reuse domain-independent planning, (ii) planning knowledge translation and interchange, and (iii) hierarchical plan merging. One way a planner can overcome knowledge deficiencies is to utilize plans, or pieces of plans, generated by other planners. In the first part of this dissertation, we present the theory and implementation of a multi-reuse planner, CBPOP, and show how it addresses the multi-reuse planning problems. In particular, we present novel approaches to retrieval and refitting, and we explore the difficult issue of when to retrieve in multi-reuse scenarios. To overcome incompatibilities between heterogeneous planning knowledge representations, the second part of this work presents a novel knowledge sharing methodology for planning systems in a framework called the Knowledge Interface (KI). The KI is used to realize peer-to-peer cooperation between heterogeneous planning systems and provide automated knowledge translation between a global common ontology specification and the individual planning systems' knowledge representations. The final part of this dissertation discusses a plan merging methodology that hierarchically merges separately generated plans based on the notion of plan fragments. The plans can be generated by the same planner/problem-solver, or by different planners. This merging mechanism performs global domain-dependent optimizations that cannot be applied to the individual plans in isolation.