Abstract
Trainable adaptive controllers (TACs) are a subset of process controllers in which much of the design is done online by means of training rather than programming. The authors show how a neural-network-based architecture may be used to implement a general-purpose TAC. An example of controlling a cart-pole system (an inverted pendulum mounted on a cart) is provided. It is found that filtering of the human-teacher training data, using a dynamic model of the teacher, significantly improves the neuromorphic TAC's performance.