[Bruce Nevin (2015.11.14.12:32 ET)]
MT: This whole question of controlling risk and uncertainty perception is something that needs to be addressed within PCT, though so far as I can see, there’s no way to slot it into the hierarchy.
RM: I don’t know about “slotting it into the hierarchy” but I think Powers has already provided a pretty nice demonstration of how to model control of risk in his “Feedback Model for Behavior: Application to a Rat Experiment” which is reprinted on pp. 47- 59 of LCS I. The rats in that experiment are required to make a certain number of bar presses in a specified interval in order to avoid getting shocked. Rats in that situation respond at a rate that is high enough to reduce the risk of getting shocked to nearly zero. So the rats are clearly controlling their risk of getting shocked, keeping it far below the level it would be if the rats did nothing.
In that article, Bill is careful to avoid saying with certitude that the rats are controlling a perception of probability or risk.
BP (p. 58): It is reasonable to suppose that the hypothesis which gives the better fit is the closer to the actual nature of qi. The present analysis suffers from the defect that the distribution curve was assumed rather than measured. If an experiment were set up to record this distribution, then it would be possible to arrive at a better definition of qi.
BP (pp. 54-55): there are many [analytical, i.e. numerical] forms similar to those chosen which serve nearly as well. The actual assumptions are few indeed: the shape for the distribution curve, and the linear proportionality assumed for the rat function.
BP (p. 59): Finally, it must be remembered that the sensory apparatus of organisms contain interpretive apparatus: the input quantity may, in fact, be a function of many sensory inputs, and may come into existence only after several stages of perceptual data processing. Even when that is the case, a feedback analysis along the lines suggested here can enable the experimenter to arrive at a reasonable approximation of the actual aspect of the environment that the organism is regulating, even when that aspect is an abstraction like density, or relative size, or a probability."
In B:CP, Bill advanced proposals as to the neuroanatomy of several kinds of perceptual functions. Here, his purpose is not anatomical, but rather polemical: his purpose is to show that “The presence of feedback makes this sort of experimentation [such as that reported in Verhave (1959)]) simply the wrong approach” (p. 57).
A major difficulty attempting to “slot into the hierarchy” a perceptual function for probability is the ubiquity of probability or risk. For any controlled perception at any level one may attempt a measure of probability of successful control.
This has a clear relation to the input function. At every level except the lowest (Intensity), the input function of a given perception combines a number of perceptual signals from lower levels. For the sake of a label, call these tributary perceptions. When there is adequate perceptual input (few or no tributary perceptions are provided by memory and imagination in absence of environmental input), control is excellent and the risk of failure to control is low. The lower the proportion of tributary perceptions that come from memory and imagination rather than from the environment, the less reliable the organism’s control through the environmental feedback function. The reason, obviously, is that control in imagination is not affected by disturbances in the environment.
So risk is inversely proportional to the richness of perceptual input from the environment, where “richness” is understood relative to the given perceptual input function. An important effect of learning (whether by problem-solving or by reorganization) is to increase the richness of environment-sourced perceptual input in the perceptual input functions affected by the learning. The rat is learning how to avoid being shocked.
BP (pp.49-50): After sufficient practice for a given setting of the interval timer and a given number[-of-presses] requirement (constant during one experiment), rats would approach some equilibrium rate of pressing. Thus a relationship was explored with the setting of the interval timer as the independent variable and the equilibrium rate of pressing as the dependent variable. Each experimental point was the average of three different four-hour averages of rate of bar-pressing. Scatter among the three determinations for a single point was on the order of one press per minute.
To grasp the degree of obfuscation of data, consider: The intervals varied between 15 seconds and 300 seconds (5 minutes). A 4-hour session of 5-minute intervals comprised 48 intervals, and 4 hours of 15-second intervals comprised 960 intervals. So between 48 and 960 rates per interval were averaged, depending on the interval length, and then for each interval length these averages were further averaged for three such 4-hour sessions. For 5-minute intervals, the rates for 144 intervals were averaged, and for 15-second intervals, rates for 2,880 intervals were averaged. Is it any wonder that for ‘control’ of an average of average rates of pressing the best estimate of qi is another statistical measure, probability?
Environmental inputs for the rat are few and slender. The rat could not perceive the requisite number of presses N as a controllable variable (8 presses in one experiment, just 1 press in the other!), and control shock by getting N presses out of the way as quickly as possible. This would not work unless it also knew when the timer reset, so that it would know to execute N presses again. Did the rat reorganize to control the equilibrium rate which just avoids being shocked? That would be analogous to adjusting one’s driving speed so as to get to all the traffic lights while they are green. But no:
BP (p. 50): The average rate of bar-pressing was always much faster than the rate actually required in order to avoid shock. the reason for this can be seen in the variations in bar-pressing rate; even with the average rate at a value fast enough to avoid shock, on some trials the random variations in rate were sufficient to delay the 8th press enough after reset of the timer to permit a shock.
Is it possible the rat occasionally slowed just enough to incur a shock in order to get more perceptual input as to the conditions under which it occurs? If you have an enemy, you want to know where it is. Is the enemy even still there? The simpler explanation is that the rats were controlling shock as well as they could while still trying to recognize some relationship between the shocks and some perceptible feature of their environment, while the actually relevant features (the timer and the required rate) were hidden from them.
Data for the process of learning culminating in an ‘average equilibrium’ rate might give some clues as to what perceptual variables the rats were controlling or trying to control. There simply is not enough data–we do not have enough perceptual input without a lot of imagined input–to control the perception that the rats were controlling a perception of probability of being shocked.
Another perceptual phenomenon that is troublesome to “slot into the hierarchy” is categorization. I have proposed for almost 20 years that categorization is a natural effect of every perceptual input function at every level, insofar any given tributary signal that the input function calls for can come from a variety of environmental sources. Any bifurcation is a member of the same category, whether that configuration is further perceived as a fork in the road, a fork or crotch in a tree, or the waist and legs of a biped. This is the basis of analogy and of important aspects of associative memory.
Bill successfully modeled Verhave’s data. Does that mean that the rat is controlling shock probability? A successful model demonstrates the fact of control, but, so far, gives only a broad outline of the neuroanatomy of control.
···
On Tue, Nov 10, 2015 at 2:38 PM, Richard Marken rsmarken@gmail.com wrote:
[From Rick Marken (2015.11.10.1140)]
Martin Taylor (2015.11.09.11.25)–
MT: I think it's a mistake to talk about "the" variable one is
controlling.
RM: Good point. But I think it is possible to focus on one controlled variable at a time, knowing that the means used to control that variable are themselves controlled variables and that the controlled variable you are focusing on is also the means by which a higher order variable is controlled. For example, we can focus on the optical variables controlled when catching a ball and ignore the fact that the movements used to control these variables are themselves controlled variables and that the catching ia also the means of controlling for some other variable, like getting a person out in a game of baseball.
MT: This whole question of controlling risk and uncertainty perception
is something that needs to be addressed within PCT, though so far as
I can see, there’s no way to slot it into the hierarchy.
RM: I don’t know about “slotting it into the hierarchy” but I think Powers has already provided a pretty nice demonstration of how to model control of risk in his “Feedback Model for Behavior: Application to a Rat Experiment” which is reprinted on pp. 47- 59 of LCS I. The rats in that experiment are required to make a certain number of bar presses in a specified interval in order to avoid getting shocked. Rats in that situation respond at a rate that is high enough to reduce the risk of getting shocked to nearly zero. So the rats are clearly controlling their risk of getting shocked, keeping it far below the level it would be if the rats did nothing. In terms of the categories in the “Behavior is control” spreadsheet the rat’s behavior in this experiment could be analyzed this way:
Behavior: Shock avoidance
Controlled Variable: Risk of getting shocked
Reference State: Zero
Means: Rate of bar pressing
Disturbance: Length of interval during which press must occur
RM: The next step is to find a model that explains this behavior – a control model, since this behavior is clearly a process of control. The most important part of the model is developing a precise definition of the controlled variable. Powers actually tried two different definitions: “the probability of getting a shock” ,p.s, and “the rate at which shock occurs”, r.s. Figure 1 shows how p.s is operationalized ; r.s is just p.s/I, where I is the duration of the interval during which a response must occur to prevent a shock.
RM: Powers doesn’t mention this in the paper but p.s and r.s can be considered two different descriptions of the perception the rats are controlling when they are controlling the “risk of getting shocked”. It turns out that p.s is a better definition of the controlled variable than r.s because assuming that p.s is controlled gives a better fit to the data than assuming that r.s is controlled (see Table 1. p 50).
RM: I think this is a very nice example of a PCT model of risk control as the control of the probability of a very unwanted event (a shock in this case but it could also be getting hit by a car while crossing the street) by acting in order to keep that probability at zero.
RM: I don’t know where the perception of a probability fits into the hierarchy; any suggestions? But I think what’s important is that this lovely “Rat Experiment” paper shows that what we see as “control of risk” can be nicely accounted for by PCT as control of the perception of the probability of unwanted events.
RM: It also shows that LCS I is a goldmine!
Best
Rick
Richard S. Marken
www.mindreadings.com
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble
–