An artificial neural network that allows a robot submarine to learn to dive and adapt to change

摘要

We show a very simple artificial neural network, namely an adaline that is able to learn to imitate very accurately the behavior of a complex non-linear dynamical system, such as a small submarine. The key to its success is the use of judiciously selected non-linear inputs for it. Such an adaline can thus provide the small submarine with an internal model of itself, which it can then use to calculate the signals that control its motion. The well-known ability of the adaline to adapt rapidly to changes endows the submarine with all the adaptive features that can be wished for in artificial systems as well as in biological systems. It learns very rapidly, by itself, to control its motion and, as we show in this article, once it has learned to dive, it performs flawlessly and can further adapt to any change in its environment or in itself. In particular, we show that it can very quickly leam to compensate for changes in the currents, the viscosity of the waters in which it moves, its own mass, its buoyancy and the maximum angle of deflection of its fins.