Using Parallelism to Scale Up Machine Learning to Large Data-Analysis Problems

俄勒冈州大卫大学埃瑟林顿9412205号，尤金，62,828-12摩尔斯。走向有效的缺省推理这是一个研究项目的第一年发现，该项目涉及一种新的基于近似的、可处理的非单调推理方法的三年研究。非单调性是大多数人类推理任务的关键部分。尽管如此，目前的非单调推理形式在一般情况下本质上是无法确定的，并且除了最具限制性的情况外，在所有情况下都是难以处理的。为了解决这一难题，将结合两种已知的、弱的技术--上下文受限推理和快速、不完全的一致性测试--来开发一种强大、易处理的近似机制。然而，仅靠这两种技术是不够的。由于一致性在一阶情况下是不可判定的，上下文受限推理本身并不能保证易处理，此外，已知的快速、不完全的一致性测试在现实复杂的知识库中通常会失败。这项研究的目的是证明两种技术的结合协同产生了一种易于处理的近似非单调推理机制，该机制克服了两种技术单独使用的局限性。这种方法给出合理结果的条件将被形式化，并将发展概率论证，表明所归纳的近似对于某些有用类型的缺省推理是合理的。形式结果将被证明，表明随着额外的计算资源的消耗，这些近似收敛到完全正确的推理。然后将研究建立和维护上下文以及处理由近似引起的错误所需的机制。最后，这些想法将被实施，并根据非平凡的知识库进行经验评估。

IRI-9412205 Etherington, David University of Oregon Eugene $62,828 - 12 mos. Toward Efficient Default Reasoning This is the first year finding of a research project involving a three year study of a novel approximation-based, approach to tractable nonmonotonic reasoning. Nonmonotonicity is a critical part of most human reasoning tasks. Despite this, current nonmonotonic reasoning formalisms are inherently undecidable in the general case, and are intractable in all but the most restrictive cases. To address this intractability, two known, weak, techniques-context-limited reasoning and fast, incomplete consistency testing-will be combined to develop a powerful, tractable, approximation mechanism. Neither of these techniques, alone, suffices, however. Since consistency is undecidable in the first-order case, context limited reasoning does not, by itself, guarantee tractability, Furthermore, known fast, incomplete consistency tests generally fail in realistically-complex knowledge bases. The goal of this research is the show that the combination of the two techniques synergistically yields a tractable approximate nonmonotonic reasoning mechanism that overcomes the limitations of either technique alone. The conditions under which this approach gives justifiable results will be formalized, and probabilistic arguments will be developed showing that the approximations induced are reasonable for certain useful types of default reasoning. Formal results will be proved showing that the approximations converge to fully-correct reasoning as additional computational resources are expended. The mechanisms necessary for building and maintaining contexts and for dealing with the errors induced by the approximation will then be studied. Finally, these ideas will be implemented and evaluated empirically against non-trivial knowledge bases.