A New Fast Inference Technique based on Knowledge-Base Compilation

为了克服当前知识系统中的知识获取瓶颈问题，人们对学习和知识获取支持进行了深入的研究。与这些方法不同，我们的研究是从基于说明性知识的知识库编译的快速推理机制的角度进行的。如果我们考虑逻辑知识或其子集，即Horn子句，知识编译就必须转换为素蕴涵，这些蕴涵都是从知识库派生的但不被其他人包含的蕴涵。然而，如果我们编译整个知识库，所需的存储空间就会变得太大。因此，我们开发了一种部分编译方法，该方法编译假想推理知识库中预期有助于提高推理效率的选定部分。此外，我们还将多级逻辑电路最小化技术与知识编译相结合，以达到更高的效率。对于假设推理的快速推理，我们还利用0-1整数规划的近似求解方法，发展了一种重要的多项式时间推理机制。

To overcome the problem of knowledge acquisition bottleneck in current knowledge systems, there have been intensive researches on learning and knowledge acquisition support. Unlike these approaches, we have carried out our research from the viewpoint that a fast inference mechanism based on knowledge-base compilation from declarative knowledge is important. If we consider logical knowledge or its subset, i.e., Horn clauses, the knowledge compilation bacomes to the transformation into prime implicates, which are all the implications derivable from the knowledge-base but not subsumed by others. However, the required memory space becomes too large if we compile the entire knowledge-base. Thus we have developed a partial compilation method which compiles the selected portions being expected to contribute the inference efficiency in a hypothetical-reasoning knowledge-base. Moreover, we have combined a multi-level logic-circuit minimization technique with the knowledge compilation to achieve higher efficiency. For the fast inference of the hypothetical reasoning, we have also developed an important polynomial-time inference mechanism employing an approximate solution method of 0-1 integer programming.