Online Adaptive Controller Based on Dynamic Evolution Strategies

Abstract

The majority of non-linear systems nowadays are controlled online using rapid PI-controllers with linear characteristics. Evolutionary algorithms are rarely used, especially for online adaptive control, due to their time complexity. This paper proposes an online adaptive controller based on a dynamic evolution strategy and attempts to overcome this performance problem. The main advantage of the evolution strategies over other gradient machine learning algorithms is that they are insensitive to becoming stuck into local optima. As a result, the proposed controller is capable of responding in real-time (sampling time between 1–5 ms) and was tested on a non-linear, single-degree-of-freedom robotic mechanism. To the extent of our knowledge, this is the first application of evolutionary algorithms in such an online control. In general, the results obtained were better than the results achieved using a traditional PI-controller.