On the Length of Dynamical System Control Programs
Fabio Bonsignorio · 2007
Examples from robotics research on bipedal walkers have shown that complex control programs integrating an explicit multibody model into the control program can be outperformed by very simple reinforced learning schemas applied to a ‘fit’ mechanical structure. This seems to show that in robotic and, in general, controlled system there are relations between the information metrics of the control programs and the system dynamics. The relations shown in this paper could help to understand the relations involved by the phase space ‘footprint’ of dynamical systems and the complexity, in the information theory sense, of their control programs, and possibly to ease, giving a quantitative guidance, the development of new more ‘intelligent’ artefacts. This is relevant to control theory and practice as the control of bipedal walkers have proven to be eased in term of complexity of control programs by the so called ‘passive walker’ approaches which leverage of the intrinsic dynamic characteristics of the walker systems. It is relevant to AI research, too, where embodiment behavior based and emergent controls are widely used in the practice of robotics, but still miss a comprehensive quantitative analysis framework. Bipedal walkers are examples of non linear mechanical dynamical systems. The equation derived here could be regarded as an hint of a general connection between control and information theory, so far not completely investigated.