[From Bruce Abbott (2017.11.30.1100 EST)]
This post continues my examination of Kenneth Craik’s unfinished book, The Mechanism of Human Action. Previously [Bruce Abbott (2017.11.28.1310 EST)] I quoted a passage in which Craik describes control systems as the mechanisms behind purposive behavior in humans and animals. This time I focus on Craik’s description of a possible mechanism that is eerily similar to reorganization as later envisioned by W. Ross Ashby in his Design for a Brain, which Bill Powers subsequently borrowed for the reorganization mechanism of HPCT.
But we have still only contemplated the design of a machine which modifies its behavior quantitatively; thus its one corrective mechanism acts more or less according to the amount of disturbance and the amount of corrective action it has already taken. We might in addition provide it with a switching mechanism which could actuate various effector systems, according to which was most successful. Something very like this is done by the uniselector switches in a telephone exchange; they have a bank of contacts with a rotary wiper; if a subscriber is sought, the wiper seeks a disengaged line. It does this by testing the line on which it rests; if this has a voltage across it (that is, is engaged) it automatically steps to the next contact and so on until it comes to one which is disengaged, where it automatically stops, completing the subscriber’s circuit. Since this behavior consists in being actuated by results to make a new, distinct contact, we may call it ‘qualitative feedback’, and explore its possibilities. For instance, by coupling a number of electrical function-networks (such as linear, square-law, integrating and differentiating, or time-delay circuits) to the different contacts of a uniselector causing it to be stepped on by the remaining error-voltage resulting from its previous state of best equilibrium, when it had reduced the error to a minimum value possible when using the function-network to which it was previously connected, it would discover for itself the most appropriate kind of response to reduce the disturbance to zero. In this simple form it would still, of course, only have a limited number of possible response types available, equal to the number of networks connected to its contacts, but we shall consider further elaborations of this system later. . . .
A further and most important consideration is the time constant of the feedback system. . . . For instance, in our qualitative-feedback switch system, a time constant in the feedback would insure that the wiper rested for some time on the contact connected to the network which had previously given the lowest mean error-voltage; it would put the device at an advantage in dealing with the next similar situation, and would prevent it from being unduly upset by momentary failure to compensate for some new disturbance; it would make its action depend on the smoothed or averaged past behavior, which is strongly suggestive of the phenomena of learning and experience.
These changes in action for which we have designed a simple machine remind one of the behavior of intact and decapitated Nereid worms, described by Loeb . The former crawl about on the glass of the aquarium and if they reach a corner finally turn round and crawl out; but the decapitated ones on reaching the corner attempt to go through the glass and die as a result of attempting vainly to go forward. One might suggest that . . . [the normal worm’s] heads contain the ‘switching’ mechanism which alone can alter their type of action after a prolonged period of failure.
The system Craik describes reminds me of the ArmControlReorg demo in LCS III, which connects its 14 joint-control systems each connected to 14 joint effectors via a matrix of weights. As the system attempts to follow a set of reference levels that should produce a certain pattern of activity, the reorganizing system examines the error signals once every 60 iterations of the simulation loop and adjusts the weights. Ultimately the weights through which each control system acts are reduced to near zero for every joint actuator except for the one that produces good control of that system’s perceptual signal. Powers noted that for reorganization to work, it must act on a much slower time-scale than the control systems it acts upon, in order to give those systems a chance to demonstrate how well they are controlling.