You are granted permission for the non-commercial reproduction, distribution, display, and performance of this technical report in any format. However, this permission is only for a period of 45 (forty-five) days from the most recent time that you verified that this technical report is still available from the Department of Computer Science of the University of Maryland at College Park under terms that include this permission. All other rights are reserved by the author(s).
Brian Kettler. Case-based Planning with a High-Performance Parallel Memory. November 1995.
Case-based planning (CBP) systems, like other case-based reasoning systems, can take advantage of previous planning experience by reusing stored cases (plans) in similar situations in the future. Advantages of CBP include speedup over planning from scratch and the ability to function with limited causal domain knowledge. ``Traditional'' CBP systems with the latter advantage typically cannot produce plans from scratch because they lack the more powerful adaptation mechanisms of generative planning systems. These ``reuse-only'' CBP systems rely on retrieving a plan from the casebase that is close to a solution plan. This requires large casebases with good coverage of the problem space and the ability to encode and match cases at fine levels of detail. Many such CBP systems, however, have fallen short of these requirements. They support only small, pre-indexed casebases. Pre-indexing constrains retrieval, as does the use of less expressive feature-based case representation schemes. The encoding and matching of detailed structural relationships in cases is not possible in such systems. These systems often adapt a single plan to the target problem using methods that are ad hoc or heuristic. CAPER is a novel, domain-independent, case-based planning system with improvements over traditional reuse-only CBP systems from its use of techniques that exploit a high-performance parallel memory of cases. CAPER takes a memory-intensive approach by making frequent use of memory during all phases of planning and by using large casebases, which can be automatically seeded. Because the parallel retrieval mechanisms scale to real-world sized casebases of thousands of plans, memory does not have to be pre-indexed and thus retrieval is more flexible. Detailed queries can be used to match cases, which are stored using an expressive, graph-structured case representation scheme. Plan adaptation in CAPER borrows techniques from generative planning, such as the use of plan validations, which capture dependencies in a plan, and plan composition. These techniques are incorporated into a reuse-only CBP framework for a more principled approach to adaptation than in many reuse-only CBP systems. CAPER can also use its flexible retrieval mechanisms and case representations to retrieve patch or substitute plans from memory. (Also cross-referenced as UMIACS-TR-95-112) University of Maryland Institute for Advanced Computer Studies, Dept. of Computer Science, Univ. of Maryland,
Scott Andrews. Brian Kettler. Kutluhan Erol. James Hendler. UM Translog: A Planning Domain for the Development and Benchmarking of. June 1995.
The last twenty years of AI planning research has discovered a wide variety of planning techniques such as state-space search, hierarchical planning, case-based planning and reactive planning. These techniques have been implemented in numerous planning systems (e.g., STRIPS, SNLP, UCPOP, NONLIN, SIPE). Initially, a number of simple toy domains have been devised to assist in the analysis and evaluation of planning systems and techniques. The most well known examples are ``Blocks World'' and ``Towers of Hanoi.'' As planning systems grow in sophistication and capabilities, however, there is a clear need for planning benchmarks with matching complexity to evaluate those new features and capabilities. UM Translog is a planning domain designed specifically for this purpose. UM Translog was inspired by the CMU Transport Logistics domain developed by Manuela Veloso. UM Translog is an order of magnitude larger in size (41 actions versus 6), number of features and types interactions. It provides a rich set of entities, attributes, actions and conditions, which can be used to specify rather complex planning problems with a variety of plan interactions. The detailed set of operators provides long plans (~40 steps) with many possible solutions to the same problem, and thus this domain can also be used to evaluate the solution quality of planning systems. The UM Translog domain has been used with the UMCP, UM Nonlin, and CaPER planning systems thus far. (Also cross-referenced as UMIACS-TR-95-69) University of Maryland Institute for Advanced Computer Studies, Dept. of Computer Science, Univ. of Maryland,
Using the Parka Parallel Knowledge Representation System (Version 3.2). Brian Kettler. William Andersen. James Hendler. Sean Luke. June 1995.
Parka is a symbolic, semantic network knowledge representation system that takes advantage of the massive parallelism of supercomputers such as the Connection Machine. The Parka language has many of the features of traditional semantic net/frame-based knowledge representation languages but also supports several kinds of rapid parallel inference mechanisms that scale to large knowledge-bases of hundreds of thousands of frames or more. Parka is intended for general-purpose use and has been used thus far to support A.I. systems for case-based reasoning and data mining. This document is a user manual for the current version of Parka, version 3.2. It describes the Parka language and presents some examples of knowledge representation using Parka. Details about the parallel algorithms, implementation, and empirical results are presented elsewhere. (Also cross-referenced as UMIACS-TR-95-68) University of Maryland Institute for Advanced Computer Studies, Dept. of Computer Science, Univ. of Maryland,
Last Generated Fri Aug 11 04:01:01 EDT 2000