FAIME: A Feature based Framework to Automatically Integrate and Improve Metaheuristics via Examples.

给出一些示例问题，FAIME将通过修改源代码来设计定制的元启发式算法。与人类设计的元启发式相比，这些算法将更好、更快、更强。初步研究表明，它们将比人类设计的元启发式算法执行得更好，因为我们正在将它们作为一个起点。它们的设计速度将比手动设计过程快。我们将能够对它们的性质做出更强有力的统计陈述。简而言之，我们将使用机器学习方法来设计元启发式算法，探索如何使用特征来进行预测，以及在给定问题的特征的情况下如何有效地设计元启发式算法。计算问题大致有三种类型：容易(易处理)、难(难处理)和不可能(无法计算)。这些问题之所以具有挑战性，纯粹是因为可能的解决方案数量之多，而尝试每一种可能性都需要不可行的时间。元启发式解决了棘手的问题，通过抽样一小部分可能性来缓解这个问题。一个缺点是不能保证性能。FAIME将检查自动设计元启发式的有效性，以及元启发式的关键性能指标，如可伸缩性、容错性、健壮性、收敛。元启发式研究中当前实践中存在的问题，我们计划直接解决。这些原则包括：-没有几个原则可以帮助设计元启发式算法，使其成为一个反复试验的过程。基于特征的方法将提供一个框架来解决这个问题。-已发表的元启发式算法的数量呈爆炸式增长，特别是受隐喻启发的算法，但几乎没有准则可供选择。基于特征的方法可以提供一些指导的基础。-元启发式通常是与它们将应用到的问题隔离设计的。FAIME将使用来自示例问题的反馈来告知设计过程。此外，手动设计元启发式算法也有其局限性。虽然它确实产生了性能更好的元启发式算法，但它并不有助于理解如何产生更好的元启发式算法。FAIME通过自动生成元启发式规则，然后使用功能来提供分类框架，为理解其设计提供基础，从而深入进行分析。这个FAIME项目将自动设计元启发式算法，例如使用由两个级别组成的工厂来设计问题，生成级别和测试级别生成和测试元启发式规则。元启发式将通过修改源代码生成，然后在一组问题上进行测试。我们将-研究元启发式的特性(可伸缩性、容错性、健壮性、收敛)。-使用一种采用现有源代码并对其进行改进的最新技术。-在并行机器上实施，这是非常适合设计过程的。-利用可计算性理论、概率理论和其他最新的数学结果来指导设计过程和基于特征的方法。我们还没有一个有效的元启发式和问题的分类方案。反映被分类实体之间关系的分类方案使分类变得清晰，并允许作出预测。使用基于特征的分类方案，我们将调查关于元启发式和问题的预测的准确性。我们的目标是能够对元启发式算法的性能做出统计声明，例如，元启发式H将在Z个时间单位内提供具有容差+/-Y单位的质量X单位的解决方案，从而提供一些性能指示以及为哪个问题选择哪个元启发式方法。

Given a number of example problems, FAIME will design metaheuristics tailored by modifying their source code. Compared to human designed metaheuristics, these will be better, faster and stronger. Pilot studies indicate they will perform better than human designed metaheuristics, as we are using them as a starting point. They will be designed faster than via the manual design process. We will able to make stronger statistical statements about their properties. In short, we will employ a machine learning approach to design metaheuristics, explore how features can be used to make predictions, and how effectively we can design metaheuristics given a problem's features. There are three broad types of computational problems; easy (tractable), hard (intractable), and impossible (incomputable). These problems are challenging purely because of the astronomical number of possible solutions and trying every possibility would require an infeasible amount of time. Metaheuristics tackle the hard problems, mitigating this issue by sampling a tiny subset of the possibilities. One drawback is that there is no guarantee of performance. FAIME will examine the efficacy of automatically designing metaheuristics, along with key performance indicators of the metaheuristics, such as scalability, tolerance, robustness, convergence.There are issues with current practices in metaheuristic research which we plan to tackle directly. These include the following:-There are few principles to assist with the design of metaheuristics, making it a trial and error process. A feature based approach will provide a framework against which to address this. -There is an explosion in the number of published metaheuristics, in particular metaphorically inspired ones, with few guidelines to select among them. A feature based approach can provide a basis offering some guidance. -Metaheuristics are often designed in isolation from the problems they will be applied to. FAIME will use feedback from the example problems to inform the design process. Also, manually designing metaheuristics has its limitations. While it does produce better performing metaheuristics, it does not contribute to an understanding of how to produce better metaheuristics. FAIME drives deeper by automatically generating metaheuristics and then using features to provide a framework for classification, providing a basis to understand their design. This FAIME project will-automatically designs metaheuristics for example problems using a factory consisting of two levels, a generating-level and a test-level which generates and tests metaheuristics. The metaheuristics will be generated by modifying their source code, and then tested on a set of problems. We will -investigate the properties of the metaheuristics (scalability, tolerance, robustness, convergence).-employ a recent technique which takes existing source code and improves it.-be implemented on parallel machines, which are ideally suited to the design process.-draw on computability theory, probability theory, and other recent mathematical results to inform the design process and feature based approach.We do not yet have an effective classification scheme for metaheuristics and problems. A classification scheme which reflects the relationships of entities being classified brings clarity and allows predictions to be made. Using the feature based classification scheme, we will investigate the accuracy of predictions we can make about metaheuristics and problems. We aim to be able to make statistical statements about the performance of metaheuristics such as, metaheuristic H will deliver a solution of quality X units, with tolerance +/- Y units, within Z time units, thus offering some indication of performance and which metaheuristic to choose for which problem.