Some approaches within AI appear to have some parallels with PCT. However, the similarities are largely superficial, and the differences identified here:
Fuzzy Logic – although Fuzzy Logic deals with continuous rather than discrete variables it is still defining transfer functions between input and output, albeit in a probabilistic or loosely defined way. It still requires definition of mapping rules and doesn’t explicitly represent a goal-based system. Perceptual control, on the other hand, is a goal seeking system and automatically adjusts its output accordingly.
Predictive Coding – Similarly to PCT, Predictive Coding does regard perception and action as closely-coupled within a hierarchical architecture. However, the descending signals are thought of as predictions of what the perception should be rather than goals to be achieved. This still requires a complex (Bayesian) mapping, in this case, between current state and predicted perception. Although this may capture some observational aspects of perceptual experience we do not see that it provides an adequate explanation of active control within dynamic behaviour.
Memory-prediction Framework (Predictive Memory) – The
memory-prediction paradigm (Jeff Hawkins, On Intelligence) concerns, like PCT, hierarchies of generic components, However, the memories are supposed to be stored values of the outputs of the system, which are then replayed when a task is repeated. It gives the example of baseball catching in terms of a fielder remembering the temporal sequence of muscle tensions used to run the path to the ball. This seems highly impractical as it would require an enormous amount of data for all the muscle tension values throughout the trajectory, and would only be valid if the environment did not change from the previous run.
Finite-State Automata – At first glance a FSA network may look similar to a PCT network, however, they are very different beasts. A FSA node consists of an initial state and an end state, and the transition between the two. Only one node in the network is ‘active’ at any time. The route through the network represents a pre-determined sequence of actions to be taken, depending upon the external inputs. In other words, it is a predictive, discrete, open-loop model of actions in constrained environments. Within a PCT network, on the other hand, all nodes are active simultaneously with each varying their outputs in order to maintain perceptual goals. The outputs that occur are not pre-determined but adapt to circumstances.
Multi-Agent Systems – Multi-agent systems are composed of multiple interacting agents that can solve problems difficult for an individual agent to solve. In this sense then PCT systems could be regarded as multi-agent systems. However, the multi-agent system paradigm does not specify the technology on which the agents are based. Generally, they are not based upon PCT, so PCT systems are a special case of multi-agent systems. A PCT hierarchy could be regarded as a multi-agent system, where each of the PCT units are interdependent agents. Also, a PCT hierarchy encapsulated within one ‘robot’ body could be regarded as an individual agent which then interacts with other independent PCT robots to form a multi-agent system.
Visual-Servoing – A crude PCT system which just uses visual feedback to control a robot system (usually arm) is very similar to visual-servoing. However, visual-servoing is generally based upon computational aspects of the scene such as the 3D pose of the robot or on the coordinates of visual features. A PCT system would not need to ‘compute’ objective aspects of the world, but would perceive more subjective relationships that lead to a goal. Although the two paradigms share some common concepts PCT takes the principle of perceptual feedback far beyond visual-servoing and applies it to all behaviour in a much more flexible and comprehensive way.
Neuromorphic Computing – Both PCT and Neuromorphic Computing presume to mimic the neuro-biological architecture of the nervous system. However, Neuromorphic Computing focuses too closely on the physical nature of neurons (with respect to spiking) and still view neuronal systems as input/output systems. This loses sight of what we see as the functional architecture of neuronal systems, that of negative feedback systems that affect and maintain the perceptual input: perceptual control.