Idealization-Based Discovery of Search Heuristics

许多计算问题可以通过在一个定义良好但巨大的候选空间中搜索来解决，只要有良好的启发式方法来指导搜索。因此，搜索启发式的自动发现将具有实用和科学的重要性。在这种方法中，启发式的生成被建模为对初始问题规范的连续转换的应用。本研究解决了扩展到更现实的应用领域的长期目标所引起的一些问题。为了完善这个模型，将进行旨在理解它何时以及为什么起作用的微扰实验研究。为了丰富生成的启发式算法的空间，将对变换进行扩展和集成，从而重新推导已知的调度和路由启发式算法，并发现新的启发式算法。为了更有效地探索这个空间，将采用新的方法来评估潜在的启发式。这项工作的预期结果包括发现启发式的自动化方法的工程进展，以及理解其潜在原理的科学进展。

Many computational problems can be solved by searching through a well-defined but huge space of candidates provided good heuristics are available to guide the search. Thus automated discovery of search heuristics would have both practical and scientific importance. In this approach, the generation of heuristics is modelled as the application of successive transformations to an initial problem specification. This research addresses a number of issues raised by the long-term goal of scaling up to more realistic application domains. To refine the model, studies of perturbation experiments aimed at understanding when and why it works will be carried out. To enrich the space of generated heuristics, extensions and integration of transformations so as to rederive known scheduling and routing heuristics and the discovery of new ones will be undertaken. To explore this space more efficiently, novel methods for evaluating potential heuristics will be employed. Expected outcomes of this work include engineering advances in automated methods for discovering heuristics, as well as scientific advances in understanding their underlying principles.