Evolved neural fields applied to the stability problem of a simple biped walking model

This paper proposes an evolved control architecture based on neural fields for a relatively complex and unstable dynamical system. The neural field model is capable of addressing goal-based planning problems and has properties, like embedding in an Euclidean space and linear stability, that potentially make it well-fitted for dynamic control tasks. The neural field control architecture is tested over the stability problem on a typical inverted-pendulum and the performance of an evolved neural field and a hand-tuned neural field is compared. The neural field controller performs well in the simulation and has a spatial representation which allows interpretation of field potentials. Also, the evolved neural field performs almost as good as the non-evolved one, is more general, and uses a different strategy to control the plant.