The Prescience of Powers

[MartinT]

In 1994, Powers presented to the first European workshop on PCT a design for what he called an “Artificial Cerebellum” (AC) that could serve as an Output Function for any control loop at any level of the hierarchy (available at http://www.pctweb.org/Powers_cerebellum.pdf. It did not matter whether the AC was usd at a peripheral perceptual control level or at the very highest (cognitive?) level. The AC compensated for the spectral characteristics of the entire control loop, thereby both speeding the output when the loop characteristic was not flat, and therefore producing a certain amount of predictability, improving control whether of actin in the evironment or imaginative thought (though Powers was not so explicit in its range of usefulness).

A few days ago I ran across a paper published 20 years later in 2014 entitled “The cerebellum for jocks and nerds alike” at https://www.frontiersin.org/articles/10.3389/fnsys.2014.00113/full. Here’s the abstract, which I think sounds as though it coud have been an abstract for Powrs’ AC presentation, apart from discussing nural spikes instea of neural current.

Historically the cerebellum has been implicated in the control of movement. However, the cerebellum’s role in non-motor functions, including cognitive and emotional processes, has also received increasing attention. Starting from the premise that the uniform architecture of the cerebellum underlies a common mode of information processing, this review examines recent electrophysiological findings on the motor signals encoded in the cerebellar cortex and then relates these signals to observations in the non-motor domain. Simple spike firing of individual Purkinje cells encodes performance errors, both predicting upcoming errors as well as providing feedback about those errors. Further, this dual temporal encoding of prediction and feedback involves a change in the sign of the simple spike modulation. Therefore, Purkinje cell simple spike firing both predicts and responds to feedback about a specific parameter, consistent with computing sensory prediction errors in which the predictions about the consequences of a motor command are compared with the feedback resulting from the motor command execution. These new findings are in contrast with the historical view that complex spikes encode errors. Evaluation of the kinematic coding in the simple spike discharge shows the same dual temporal encoding, suggesting this is a common mode of signal processing in the cerebellar cortex. Decoding analyses show the considerable accuracy of the predictions provided by Purkinje cells across a range of times. Further, individual Purkinje cells encode linearly and independently a multitude of signals, both kinematic and performance errors. Therefore, the cerebellar cortex’s capacity to make associations across different sensory, motor and non-motor signals is large. The results from studying how Purkinje cells encode movement signals suggest that the cerebellar cortex circuitry can support associative learning, sequencing, working memory, and forward internal models in non-motor domains.

[rsmarken]

@MartinT

In 1994, Powers presented to the first European workshop on PCT a design for what he called an “Artificial Cerebellum” (AC) that could serve as an Output Function for any control loop at any level of the hierarchy

I think Powers’ overroad his own prescience when he realized that the two level hierarchy could behave adaptively sans the artificial cerebellum. Note that Figure 4 in the Powers (1994) An “Artificial Cerebellum” Adaptive Stabilization of a Control System paper from the Wales conference (great conference!!!) is exactly equivalent to Figure 5.2 in Powers (2008) Living Control Systems III. The only difference is the A.C.s (artificial cerebellum’s) in Figure 4. Comparison of Figure 5.5 in the 2008 paper to Figure 6 in the 1994 paper suggests that the A. C. either provided no improvement or made things slightly worse. There may be a role for an A.C.somewhere in the control processes of living control systems but it doesn’t seem to be needed to account for the apparent “adaptation” the occurs when when you change the load or spring constant when controlling the position of a mass on a spring. Just a simple, two level control organization.

Best

Rick

PS. Don’t you and your friends want to attend the IAPCT conference? It’s hard to beat the price;-)

[MartinT]

I’m not going to argue with you on this, because I think you are essentially correct. Powers did set the AC aside in favour of a reorganizing 2-level control hierarchy.

But that wasn’t the point of my posting. My point was that the abstract of a 2014 paper describing a complex bit of physiology could well have served as the abstract for the functional design that Powers had described 20 years earlier (apart for the substitution of neural spikes for neural current).

It seems to me that when a design from PCT turns out to correspond to a complex bit of physiology, the PCT design deserves to be considered seriously, even when another PCT design (in this case a reorganizing hierarchy) will do essentially the same job.

There’s no reason to suppose that only one of the equivalent mechanisms functions in the physiological brain, and in this case it might even be that the cerebellum is where the actual non-conscious reorganizing hierarchy resides. The two figures might actually be the same figure, with the difference that one is conceptually linked to a particular piece of brain physiology while the other isn’t.

[rsmarken]

MT: Powers did set the AC aside in favour of a reorganizing 2-level control hierarchy.

RM: There was no reorganizing involved in the 2-level control hierarchy in LCS III. That was the beauty of it; adaptation sans adaptation.

[MartinT]

I guess I should thank Rick for reminding me that there probably are PCT neophytes lurking in this forum. For their benefit I should explain the point of his comment, that almost no reorganization is done during the LCS III demo in question. Reorganization changes only the structure and parameters of the control hierarchy, and does not take part in actual controlling. It’s ordinarily much slower.

In this case, Bill Powers himself took the part of the reorganizing system, building the structure we see in LCS III Figure 5.2 before starting the demo, and altering the two gain parameters as part of the demo. My own comment to which Rick correctly replied was, I assume, clear to anyone familiar with PCT, but probably not to a novice, so I again thank Rick for the clarification.

[rsmarken]

Martin Taylor says:
MT: I guess I should thank Rick for reminding me that there probably are PCT neophytes lurking in this forum. For their benefit I should explain the point of his comment, that almost no reorganization is done during the LCS III demo in question.

RM: I think there is a good lesson in that demo for everyone, not just neophytes. The demo in question is the two level control hierarchy (Figure 5-2) described in Chapter 5 of LCS III. That chapter is titled “Non-Adaptive Adaptive Control” because it describes another behavioral illusion – the illusion that a control system changes its own characteristics in order to adapt to a changed environment when, in fact, it does not change at all.

RM: The characteristics of the system that appear to be changing are dynamic characteristics that presumably change in response to changes in the dynamic characteristics of the variable being controlled (in this case, the position of a mass on a spring). Changes in system characteristics are thought to be reflected in what are called Bode plots such as these:

RM: A Bode plot shows the frequency response of the system to a disturbance to the controlled variable. The frequency response presumably reflects control characteristics of the system doing the controlling. So the three Bode plots above presumably reflect a system with three different control characteristics (mainly characteristics of the system’s output function). These different control characteristics are seen for the same system controlling the position of a mass (the load) with different dynamic characteristics, called proportional, single integral and double integral.

RM: Engineering psychologists have interpreted these differences in control characteristics as a result of adaptation to the different dynamic characteristics of the load. In PCT we would call this kind of adaptation “reorganization” because it presumably reflects a change in the parameters of the control system itself. But in fact these different Bode plots were produced by the same two-level control system with all the same parameters ((Figure 5-2). That is, taking these different Bode plots as reflecting changes in control characteristics resulting from adaptation to a changing environment is another example of a behavioral illusion. Here’s how Bill puts it in LCS III:

WTP: This will come as a surprise to many engineering psychologists, because about 40 years ago several of their ancestors (McRues and Jex, 1967) determined experimentally that human control characteristics do change when the characteristics of the load change from leading to proportional to lagging. There is no implication here that their measurements were in error; they were correct but what they measured was probably an illusion. (Powers, LCS III, p.84-85) Italics are Bill’s; bold emphasis mine.

MT: In this case, Bill Powers himself took the part of the reorganizing system, building the structure we see in LCS III Figure 5.2 before starting the demo, and altering the two gain parameters as part of the demo.

RM: There was no reorganization (adaptation) involved in the production of this demo; the three different Bode plots shown above were produce by a system with the same control parameters. The only thing that differed in the three cases was the dynamic characteristic of the load. Taking the different plots as evidence that reorganization has occurred is a mistake – another example of a behavioral illusion.

RM: As Bill points out at the end of Ch. 5, “this demonstration is not to do away with adaptation [reorganization – RM]; it is to show that negative feedback control systems can be designed very simply so that they operate over a wide range of environmental properties without requiring any adaptation at all”. (LCS III, p. 94). As with any behavioral illusion, you have to use modeling – which involves knowing the variable(s) being controlled – to tell whether an apparent change in system characteristics is a result of adaptation or (as in this case) not.

Best

Rick

[MartinT]

In effect, Rick is denying that Bill designed and implemented the control structure (that he did not reorganize the structure) and that he did not direct the student to arbitrarily change the two gain parameters (that he or the student did not reorganize the parameters).

I do not think that the particular hierarchic control structure emerged by magic into the pages of the book and into the program that describes it, so in this I must dispute Rick’s claim that There was no reorganization (adaptation) involved in the production of this demo; [As a side note, one might observe that this structure contradicts the ordering of the low levels in the “standard” eleven-level hierarchy].

Other than that, I see no need to comment further in this thread, which was started simply to show how by considering a way of performing a function, Bill produced a structure (the Artificial Cerebellum) which very closely matches what neurophysiologists described as existing in the physiological cerebellum. I consider that to be a fine tribute to his prescience, whether or not he followed the AC idea in later work.

How multi-level hierarchic control structures can or do function is quite off-topic, however interesting might be the fact that in a particular task the 1-level AC and a 2-level hierarchy produce almost exactly the same performance. It would be interesting to see how far this mimicry extends, and which form of adaptation, AC or hierarchic reorganization works faster or better. Following that trail might lead to interesting results relating to reorganization and consciousness, but it’s not the topic of this thread.

[rsmarken]

Hi Martin

MT: In effect, Rick is denying that Bill designed and implemented the control structure (that he did not reorganize the structure) and that he did not direct the student to arbitrarily change the two gain parameters (that he or the student did not reorganize the parameters).

RM: Bill certainly designed the two level control structure, so, in a sense, he acted as the reorganizing system that developed that control hierarchy. My point was simply that, because there is there is no reorganization involved in the two-level model – the apparent adaptation to a change in the dynamics of a controlled variable does not necessarily involve any reorganization at all. A simple, two level control organization achieves, sans reorganization, what a one level control organization achieves with it (in the form of the artificial cerebellum).

MT: I do not think that the particular hierarchic control structure emerged by magic into the pages of the book and into the program that describes it, so in this I must dispute Rick’s claim that There was no reorganization (adaptation) involved in the production of this demo;

RM: I suppose my statement could be taken as being ambiguous. Instead of saying that there was no reorganization involved in the production of the demo I should have said that there was no reorganization involved in the production of the apparent adaptation seen in the behavior of the two-level control hierarchy when the dynamics of the controlled variable (position of the load) are changed.

MT: [As a side note, one might observe that this structure contradicts the ordering of the low levels in the “standard” eleven-level hierarchy].

RM: Good point. And all the more reason to dedicate research to testing Powers’ theory of hierarchical control; the types of perceptual variable controlled at each level of the control hierarchy is a hypothesis, not received truth.

MT: …Bill produced a structure (the Artificial Cerebellum) which very closely matches what neurophysiologists described as existing in the physiological cerebellum. I consider that to be a fine tribute to his prescience, whether or not he followed the AC idea in later work.

RM: I think it shows the problems involved in going from physiology to behavioral model rather than the other way around. As clever as the AC is, the fact that it apparently matches some neurophysiology may be misleading. The AC was developed without the physiology in mind. It was built to explain the observed ability of organisms to produce stable control in an environment with changing dynamic characteristics. Bill later discovered that this same apparent adaptation could be accomplished by a two level hierarchy. So now we have two models of the same phenomenon. I think this provides an opportunity to compare the models against actual behavioral data. Once the best model is found it would be time to do the neurophysiological work to see how the process is implemented in the nervous system.

MT: How multi-level hierarchic control structures can or do function is quite off-topic, however interesting might be the fact that in a particular task the 1-level AC and a 2-level hierarchy produce almost exactly the same performance. It would be interesting to see how far this mimicry extends, and which form of adaptation, AC or hierarchic reorganization works faster or better. Following that trail might lead to interesting results relating to reorganization and consciousness, but it’s not the topic of this thread.

RM: I think following that trail would lead to a dead end because you are not comparing the model behavior to actual behavior. What would be interesting is to do what I suggested above: see which model – AC or two-level hierarchy (there is no reorganization involved in the two level hierarchy model) – best accounts for the data.

Best

Rick