Accelerating Evolutionary Design Exploration with Predictive and Generative Models

摘要: Optimization plays an essential role in industrial design, but is not limited to minimization of a simple function, such as cost or strength. These tools are also used conceptual phases, better understand what possible. To support this exploration we focus on Quality Diversity (QD) algorithms, which produce sets varied, high performing solutions. techniques often require the evaluation millions solutions -- making them impractical design cases. In thesis propose methods radically improve data-efficiency QD with machine learning, enabling its application design. our first contribution, develop method modeling performance evolved neural networks for control and The structures these grow change, difficult model new able estimate their based heredity, improving by several times. second contribution combine model-based optimization MAP-Elites, algorithm. A created from known designs, MAP-Elites creates set designs using approximation. subset evaluated model, process repeats. We show that approach improves efficiency orders magnitude. Our third integrates generative models into learn domain specific encodings. variational autoencoder trained produced capturing common “recipe” performance. This learned encoding can then be reused other algorithms rapid optimization, including MAP-Elites. Throughout thesis, though vision examine applications fields, robotics. advances exclusive serve foundational work integration learning.