Perceptual control theory (PCT) is a psychological theory of animal and human behavior. PCT postulates that an organism’s behavior is a means of controlling its perceptions. The model is based on the principles of negative feedback . It is to some extent an application of the ideas used in the engineering discipline of control theory to the modeling of the human mind and behavior.
PCT postulates that layers of control systems, which have access to a metric to optimize and some set of policies or actions, can maintain balancing-acts for difficult, high-abstraction things without developing any explicit model for how those actions relate to the metric being tracked. The brain is postulated to be one of these multi-layered PCTs.
Physical movements are a favorite case-study, since they’re relatively easy to break down into this these sorts of layered control theory sub-problems. An important subtlety is that the control systems are optimizing for the perception of a state, rather than for a concrete environmental state itself.
Actions and behaviors develop because they do something to reduce the mismatch between internal perception, and the stimulus readings received.
It’s unclear whether PCT is a valid theory. It doesn’t significantly constrain the space of possible minds that could be built from it, and the advocates of the theory on the blog were unable to make a clear case for it. Experimental results for its quality as an algorithm seemed lackluster; see these critical comments about the paper version of this technical report, which claim that the correct results may have been achieved more through parameter-fitting than PCT.
Some anecdotally found it more useful for explaining bugs in human behavior, than for modeling what would be ideal behavior.
Under-Characterized Information Storage
This seems to be a common category of complaints about Perceptual Control Theory.
This blog post called out that PCT “has no theory of information or how that information comes to be made,” and this post grappled with a similar problem: struggling to find a place for implicit models, priors, and updates when working with a PCT framework. (This comment may have made a case for at least some embedded implicit priors.)
His personal-blog thoughts on the topic here