| Advancements in the field of machine learning has made a model-free approach for nonlinear control of dynamical systems more viable. Traditionally, the controller design is based on the analysis of the system model. In practice, however, it might not be possible to estimate a system model that truly reflects the complex behavior of the real system. A model-free controller self-learns the required control decisions by applying machine learning techniques, avoiding the need for estimation and analytical design. In this thesis, a parameterized dynamical system known as Dynamic Movement Primitive (DMP) is used as a feedforward model-free controller. An advanced, nature-inspired, evolutionary machine learning algorithm called Covariance Matrix Adaption Evolution Strategy (CMA-ES) was used to self-learn the control decisions. It was demonstrated through computer simulated experiments that such an evolutionary model-free controller could successfully learn to accomplish the difficult task of swinging up a double inverted pendulum, motivating further research. |
