Taming the fluidic pinball with artificial intelligence control
Abstract
The aim of this work is to develop a generic control strategy for nonlinear dynamics. This strategy is based on genetic programming, a machine learning technique for regression problems, that maps the sensor signals to the actuators in a unsupervised manner.
It's a biological inspired method mimicking Darwin's natural selection: through and evolution process it derives a control law minimizing a given objective.
Genetic programming has been applied to a DNS of a 2D fluidic mechanic system, the fluidic pinball.
Several search spaces including control laws built from periodic functions, sensor signals and time-delay sensor signals have been explored.
For the fluidic pinball genetic programming managed a 46% net drag saving, outperforming by 3.3% the best open-loop control law found with a parametric study.
Our contribution has been the acceleration of the learning process by avoiding the evaluations of redundant control laws, thus improving the learning rate by a factor 3.