An Exploration-Estimation Algorithm for Synthesis and Analysis of Engineering Systems Using Minimal Physical Testing
Hod Lipson, Josh Bongard · 2004
We describe a new general algorithm for the automated design, analysis and repair of nonlinear physical systems. The process iterates a two-phase exploration-estimation cycle. The exploratory phase seeks a new improvement or test to perform to the system based on some initial internal model. The estimation phase performs the suggested operation and observes the outcomes; it then improves the internal model so as to explain all observations so far. This process relies on very few, targeted, and carefully planned interactions with the physical systems. We describe an implementation of this method using two evolutionary algorithms, where the exploratory phase uses a simulator to evolve improvements or tests, and the estimation phase uses observations to evolve the simulator itself. We demonstrate this algorithm for analysis, design and repair of electromechanical systems.