ABSTRACT
A neural network implementation is proposed, such a network uses the least information possible, and is able to learn on-line, adapting its weights in a continuous way using reinforcement learning. At the beginning, the network behaves as a reflex conditioned system, after being vaguely initialized, yielding a rough control signal very similar to bang-bang controllers, but as learning progresses, the response is gradually smoothed. A similar process can be observed in animal behaviour during learning. As an application, the network is used to control a chemical batch reactor. This kind of reactors are difficult to control: they are highly non-linear, they have time varying parameters and no stable state, and the kind of chemical processes and operating conditions change very often.
