Perceptual control networks are a special kind of neural network based on Perceptual Control Theory, as pioneered by W.T. Powers.
The idea that living systems work by keeping certain variables in their environment at fixed points has become a central notion in modern biology. Control of a variable means acting on the world to keep that variable within certain predefined limits. A system that works in a sufficiently restricted or predictable environment may be able to do so without looking at the result of what it is doing. This is known as an open loop system: it takes a stimulus in and computes the appropriate response to bring the controlled variable to a desired state. This kind of system cannot function properly in an ever-changing, complex world.
A system that functions in a world full of complex disturbances (a living organism, for example) must look at what it is doing while it is doing it. The system's outputs correct error continuously, receive feedback on the value of the controlled variable, and continuously compute new output. Of course, the system is not really seeing the result of its actions at the instant it acts. Some delay is unavoidable, but should be as low as possible. This kind of control system cannot be understood as simple cause-effect or stimulus-response. Both input and output are stimulus and response, cause and effect. In other words, they function in an interdependent closed loop.
An important insight in the Powers version of control theory is the issue of what is being controlled. Wiener said that it was an environmental variable under control--something out there in the world. Ashby said it was the organism's essential variables--basic physiological states important for survival. Finally, Powers said that it was the organism's perception that was under control. In this formulation, controllers act to match their perceptual input with a reference passed down from higher level controllers. Under this theory, the human brain is a large hierarchy of simple controllers, each controlling a single scalar variable, but together appearing to control complex multi-dimensional variables at many different levels. The higher you go up the hierarchy, the longer the feedback delays become. Ultimately, at the very highest level, we are all controlling for Ashby's essential physiological variables. But at intermediate levels in the hierarchy, we control for all kinds of more "psychological" references, such as perceiving our bank balances to be high, or perceiving a mate to be attractive.
NTT Systems built two Control Builder simulators for Dr. Martin Taylor at DCIEM for doing research on perceptual control networks. We have performed preliminary experiments at getting such networks to track a simplified "toy" speech system. We have also built a tool for Layered Protocols, a design paradigm for user-interfaces, based on a perceptual control approach.
NTT Systems has a literature survey, available on request, of perceptual control and related fields.