Non-Darwinian Evolutionary Computation: Guiding Evolution by Machine Learning

这是一个为期三年的连续奖励的第一年资助。 该项目涉及一种新的进化计算方法，称为可学习进化模型（LEM），它从“智力进化”中汲取灵感-人类思想，设计，人工制品，组织等的进化。与传统的达尔文型进化算法不同，其中新的候选解决方案是通过半随机突变和/或重组创建的，LEM通过假设形成（通过机器学习）和假设实例化（通过演绎）生成新的解决方案。 初步研究表明，LEM可以将进化过程加速两个或更多个数量级，超过达尔文型算法（就出生或适应性评估的数量而言），并且因此在适应性评估耗时和/或昂贵的应用领域中似乎特别有利，例如工程优化问题、进化设计、流体动力学、药物开发、可进化硬件，本研究将探讨LEM的能力，局限性和复杂的优化和工程设计任务的可扩展性。 PI将开发初始方法的扩展，并将在选定的现实问题领域中对LEM进行实验测试。

This is the first year funding of a three year continuing award. This project concerns a new approach to evolutionary computation, called Learnable Evolution Model (LEM), which draws inspiration from "intellectual evolution" - an evolution of human ideas, designs, artifacts, organizations, etc. Unlike conventional Darwinian-type evolutionary algorithms, in which new candidate solutions are created by semi-random mutations and/or recombinations, LEM generates new solutions by hypothesis formation (via machine learning) and hypothesis instantiation (via deduction). Pilot studies have shown that LEM can accelerate evolutionary processes by two or more orders of magnitude over Darwininan-type algorithms (in terms of the number of births or fitness evaluations), and thus appears to be particularly advantageous in application domains in which fitness evaluations are time-consuming and/or costly, such as engineering optimization problems, evolutionary design, fluid dynamics, drug development, evolvable hardware, geological inversion, etc. This research will explore LEM's capabilities, limitations, and scalability to complex optimization and engineering design tasks. The PI will develop extensions of the initial methodology, and will experimentally test LEM in selected real-world problem domains.