SGER: Learnable Evolution: Speeding up Evolutionary Computation by Inductive Learning

IIS-9904078乔治梅森大学Ryszard Michalski$49,751-12 mos SGER：可学习进化-通过归纳学习加速进化计算在这项探索性研究中，PI将继续研究一种新的进化计算方法，他称之为可学习进化模型(LEM)。与作为当前进化计算方法基础的达尔文进化模型不同，LEM涉及一个归纳学习过程，该过程寻找为什么种群中的某些个体在执行指定类别的任务时优于其他个体的原因。这些被表述为归纳假设的原因被用来在进化学习过程的每一步创造新一代的个体。在一项已经完成的初步研究中，将LEM算法与两种标准遗传算法在解决不同类型的优化问题方面进行了比较。这些问题已经被进化计算界广泛用于测试遗传算法。在这些实验中，LEM的性能远远超过了遗传算法，经常实现两到三个数量级的加速。在某些情况下，LEM找到了最优解，而研究中使用的遗传算法离最优解还很远。根据这一奖项，PI将对LEM方法进行理论和实验研究，在不同类型的问题上测试它，特别是在非常高维和噪音的优化问题上，以及系统设计问题上。他将研究它的局限性，并确定通过与他之前在建设性归纳(在NSF支持下开发)相关的想法和方法来增强它的方法。所提出的研究可能会导致进化计算和机器学习领域的重大进展，这些领域可以应用于超复杂优化问题、工程系统设计、大型数据库中的知识发现和自我完善的智能代理等领域。

IIS-9904078Ryszard Michalski, George Mason University$49,751 - 12 mos SGER: Learnable Evolution - Speeding up Evolutionary Computation byInductive LearningIn this exploratory research the PI will continue work on a new approach to evolutionary computation that he calls Learnable Evolution Model (LEM). In contrast to the Darwinian evolution model that underlies current methods of evolutionary computation, LEM involves an inductive learning process that seeks "reasons" why certain individuals in a population are superior to others in performing a designated class of tasks. These reasons, expressed as inductive hypotheses, are used to create a new generation of individuals at each step of an evolutionary learning process. In an already completed pilot study, LEM has been compared to two standard genetic algorithms in solving different kinds of optimization problems. These problems have been widely used by the evolutionary computation community for testing genetic algorithms. In these experiments, LEM outperformed the genetic algorithms by a wide margin, frequently achieving a speed-up of two or three orders of magnitude. In some cases, LEM found the optimal solution, while genetic algorithms used in the study were still far away from it. Under this award, the PI will conduct theoretical and experimental studies of the LEM approach, testing it on different types of problems, in particular, on very high dimensionality and noisy optimization problems, and system design problems. He will study its limitations, and determine ways of enhancing it by ideas and methods related to his prior work on constructive induction (developed under NSF support). The proposed research may potentially lead to major advances in the fields evolutionary computation and machine learning, which can be applied to a variety of areas including ultra-complex optimization problems, design of engineering systems, knowledge discovery in very large databases, and self-improving intelligent agents.