Any comments on this detailed report folks?
From Fred Nickols (2017.01.28.0645 ET)]
I think it’s in the conclusion, Warren; namely, the inability of the algorithm-based, plan-and-control approach to cope with unforeseen, nonstandard circumstances. In PCT-talk, that would be “disturbances.”
Fred
···
From: Warren Mansell [mailto:wmansell@gmail.com]
Sent: Saturday, January 28, 2017 12:18 AM
To: csgnet@lists.illinois.edu
Subject: It’s not PCT but it works… self-driving cars…
Any comments on this detailed report folks?
[From Bruce Abbott (2017.01.28.0855 EST)]
Fred Nickols (2017.01.28.0645 ET) –
FN: I think it’s in the conclusion, Warren; namely, the inability of the algorithm-based, plan-and-control approach to cope with unforeseen, nonstandard circumstances. In PCT-talk, that would be “disturbances.”
BA: The article’s publication date is 2011. Since then, self-driving vehicles produced by private companies (e.g., Tesla, Google, Uber) have logged hundreds of thousands of miles on public roads in all sorts of environments (urban and rural) and seem to be performing at least as safely as human drivers. Apparently a number of the important problems mentioned in the article have been solved, highlighting the very rapid progress that has occurred in this area between 2011 and the present. It does not appear that the current crop of autonomous vehicles come equipped with hundreds of thousands of dollars-worth of sensors and computational power as was the case for Stanford’s test vehicle. Consequently I would not take this article as the last word in how safe autonomous driving is being accomplished by current vehicles.
BA: Fred, you raise the standard PCT complaint against the computational approach described in the article, but I think we have to be careful to distinguish what sorts of disturbances the vehicle can and cannot handle well. The bottom-level systems evidently are PID controllers (plus feedforward) for steering angle, braking, and speed control. These control systems can handle disturbances to these variables. Similar systems are at work in some current production automobiles – adaptive cruise control and systems that vary individual wheel braking and engine power to maintain control during loss of traction in turns or when stopping. Where the Stanford vehicle had trouble mainly relates to its ability to detect objects that may pose a hazard within cluttered or nonstandard environments such as construction zones and its ability to detect and correctly interpret road signs. There are no PCT models that can do as well; in fact, there are no PCT models that can do it at all, because, so far as I am aware, there is no one in PCT working on developing such models. At its present stage of development, PCT only offers a speculative hierarchy of perceptual systems and no models of any specific perceptual system that could be tested against alternative models. When developing models of, say, tracking, we simply assume the presence of a perceptual system that somehow identifies cursors on a screen and determines the distance between them. We offer no models that actually perform this function.
Bruce
···
From: Warren Mansell [mailto:wmansell@gmail.com]
Sent: Saturday, January 28, 2017 12:18 AM
To: csgnet@lists.illinois.edu
Subject: It’s not PCT but it works… self-driving cars…
Any comments on this detailed report folks?
[From Rick Marken (2017.01.28.1025)]
WM: Any comments on this detailed report folks?
Â
https://www.cs.cmu.edu/~zkolter/pubs/levinson-iv2011.pdf
RM: This is not PCT because PCT is control theory applied to understanding living organisms. What they are describing in this article is control theory applied to the engineering of self-driving cars. Actually, they are applying Modern Control Theory (MCT), which involves a lot of prediction. In LCS III (The Fact of Control) Powers explained why MCT is both unnecessary and unfeasible as a model of living organisms (see pp. 2-6).Â
RM: But even though MCT is implausible as a control organization that could be implemented in the nervous system of living systems, it is still a control organization that controls input variables and therefore acts to produce intended results in the the face of variable disturbances. That is, it can drive around in the real world – the world of continuously and unpredictably varying disturbances – fairly successfully. This is because the MCT architecture, which consists of predictive output-generating algorithms, is still a control of input system (as are all control organizations) because these outputs (accelerations, wheel angle variations, etc) have feedback effects on the inputs (perceptions) on which these predicted outputs are based. So the designers of these systems might conceive of them as working because they are producing the right outputs based on their super snazzy predictive algorithms. But they actually work because these outputs are being produced in a closed negative feedback loop that keeps the inputs under control.Â
RM: The situation is like that of Braitenberg Vehicles, which are little cars that can control for following illuminated paths, for example. The Vehicles are describes as S-R devices because they consist of light sensors hooked up to motors (output generators) that turn the wheels. But since these are actually vehicles moving in the real world, the sensor inputs exist in a closed-loop; sensor inputs cause motor outputs at the same time as the motor outputs are causing the sensor inputs. When the sensor inputs are appropriately connected to motor outputs, the feedback effects of outputs on inputs is negative and the sensor inputs will be controlled at some fixed reference value (such as zero, unless there is some offset built into the circuit).Â
RM: PCT could probably help the self-driving vehicle engineers design more efficient and effective architectures for their control systems. I think that’s what Rupert Young is doing with his Perceptual Robots project (http://www.perceptualrobots.com/).Â
···
 Bruce Abbott (2017.01.28.0855 EST)
BA: Fred, you raise the standard PCT complaint against the computational approach described in the article, but I think we have to be careful to distinguish what sorts of disturbances the vehicle can and cannot handle well.Â
RM: The “standard PCT complaint” against the computational (that is, the predictive control or MCT) approach is that it’s not neurologically or computationally feasible. But, as I said above, it works as well as it does in the case of self-driving vehicles because the algorithm is implemented in a system that is actually acting in the real world.Â
RM: I think what you dismissively refer to as the “standard PCT complaint” is the one that would be expected to raise the hackles of a conventional psychologist. That “standard PCT complaint” is about the causal model of behavior that is the basis of conventional “scientific” psychology; the complaint is about the fact that this model cannot account for the fact that consistent behavioral results are produced in the the face of unpredictable (and often undetectable) disturbances.Â
RM: This “standard PCT complaint” is the basis for the other “standard PCT complaint” that was forthrightly stated by Bill in his Foreword to my book MIND READINGS: “if the phenomenon you see here [control–RM] really works as this model [PCT–RM] shows it to work, then a whole segment of the scientific literature needs to be deposited in the wastebasket”. And we couldn’t have that, now, could we?
BestÂ
Rick
–
Richard S. MarkenÂ
"Perfection is achieved not when you have nothing more to add, but when you
have nothing left to take away.�
                --Antoine de Saint-Exupery
[From Bruce Abbott (2017.01.28.1610 EST)]
Rick Marken (2017.01.28.1025) –
Bruce Abbott (2017.01.28.0855 EST)
BA: Fred, you raise the standard PCT complaint against the computational approach described in the article, but I think we have to be careful to distinguish what sorts of disturbances the vehicle can and cannot handle well.
RM: The “standard PCT complaint” against the computational (that is, the predictive control or MCT) approach is that it’s not neurologically or computationally feasible. But, as I said above, it works as well as it does in the case of self-driving vehicles because the algorithm is implemented in a system that is actually acting in the real world.
BA: Why sound like you are disagreeing with me when you are merely identifying a different complaint, valid but irrelevant to the issue I was addressing?
BA: The “standard PCT complaint� that Fred raised was that the computational plan-compute-execute approach is unable to deal effectively with the effects of unpredictable disturbances. Are you saying that this is not a standard PCT complaint against MCT?
BA: I noted that the sorts of disturbances Fred probably had in mind are actually handled by the vehicle’s low-level controllers – the ones that compare the vehicle’™s position on the road to its reference position, compare the distance to the vehicle or other obstruction ahead to the reference minimum distance, and so on. The difficulty with handling unpredictable or unfamiliar environments mentioned at the end of the article relates more to a lack of sufficiently good perceptual algorithms than to a lack of feedback.
RM: I think what you dismissively refer to as the “standard PCT complaint” is the one that would be expected to raise the hackles of a conventional psychologist. That “standard PCT complaint” is about the causal model of behavior that is the basis of conventional “scientific” psychology; the complaint is about the fact that this model cannot account for the fact that consistent behavioral results are produced in the the face of unpredictable (and often undetectable) disturbances.
BA: Dismissively? Why the pejorative tone? I don’t think I was dismissive of anything.
RM: This “standard PCT complaint” is the basis for the other “standard PCT complaint” that was forthrightly stated by Bill in his Foreword to my book MIND READINGS: “if the phenomenon you see here [control–RM] really works as this model [PCT–RM] shows it to work, then a whole segment of the scientific literature needs to be deposited in the wastebasket”. And we couldn’t have that, now, could we?
BA: We who? Me?
Bruce
[From Rick Marken (2017.01.28.1750)]
···
 Bruce Abbott (2017.01.28.1610 EST)–
BA: Fred, you raise the standard PCT complaint against the computational approach described in the article, but I think we have to be careful to distinguish what sorts of disturbances the vehicle can and cannot handle well.Â
RM: The “standard PCT complaint” against the computational (that is, the predictive control or MCT) approach is that it’s not neurologically or computationally feasible.Â
BA: Why sound like you are disagreeing with me when you are merely identifying a different complaint, valid but irrelevant to the issue I was addressing?
RM: You are absolutely right. I am sorry. My only (lame) excuse is that  I read “standard PCT complaint” as “PCT people are always kvetching about the minor fact that open loop systems can’t deal with disturbances”; that is, I heard it as kind of dismissing the PCT point about the inability of open-loop systems to produce consistent results in a disturbance prone world as much ado about, if not nothing, at least “not much”. So I wanted to make sure it was understood that this observation is the central “raison d’être” for PCT: the fact that open-loop systems can’t produce consistent results in a disturbance prone world means that they can’t control and, as Bill showed over and over again, what we see as “behavior” is a process of control.Â
RM: Unfortunately, perhaps, I started out by noting that what you call a “computational” system can control when implemented in an actual behaving device whose outputs have an effect on the inputs that are also the cause of those outputs. Though these systems have “feedforward” (predictive) components they are still negative feedback control systems. So these systems can produce consistent results in a disturbance prone world; the PCT “complaint” about these systems, as I said, is just that their implementation in a nervous systems seems unfeasible.Â
RM: Then I went on to say that the “standard PCT complaint” that you mention is really only applicable to computational models of the behavior of living systems, and only when these models are conceived of as open-loop (as in the general linear model used in the statistical analysis of behavioral data).Â
RM: So, again, I am sorry if I misconstrued your intent when you said “standard PCT complaint”.Â
BA: The “standard PCT complaint� that Fred raised was that the computational plan-compute-execute approach is unable to deal effectively with the effects of unpredictable disturbances. Are you saying that this is not a standard PCT complaint against MCT?
RM: It is, but only when MCT is used as an “open-loop” model of behavior. MCT implemented in real systems can deal with disturbances.Â
BA: I noted that the sorts of disturbances Fred probably had in mind are actually handled by the vehicle’s low-level controllers – the ones that compare the vehicle’s position on the road to its reference position, compare the distance to the vehicle or other obstruction ahead to the reference minimum distance, and so on.Â
RM: The low level controllers themselves are very likely to be MCT controllers. An MCT controller might actually do better (and probably no worse) than a simple control loop (as Martin has noted). So the PCT argument against such controllers is that they are far too computationally intensive to be implemented in a nervous system. That’s he argument made by Bill on pp. 2-6 of LCS III.
BA: The difficulty with handling unpredictable or unfamiliar environments mentioned at the end of the article relates more to a lack of sufficiently good perceptual algorithms than to a lack of feedback.
RM: I agree. The problem of making “self-driving” cars more like real drivers is getting them to be able to perceive the higher level perceptual variables that real drivers are controlling – principle perceptions like " drive defensively" or “drive cooperatively” (ie, defer to other drivers when this would achieve the higher level  goal of keeping traffic flowing).
Â
RM: I think what you dismissively refer to as the “standard PCT complaint” is the one that would be expected to raise the hackles of a conventional psychologist.Â
BA: Dismissively? Why the pejorative tone? I don’t think I was dismissive of anything.
RM: I believe you. It was I who was perceiving the phrase as pejorative. Again, I apologize for taking umbrage inappropriately.Â
RM: This “standard PCT complaint” is the basis for the other “standard PCT complaint” that was forthrightly stated by Bill in his Foreword to my book MIND READINGS: “if the phenomenon you see here [control–RM] really works as this model [PCT–RM] shows it to work, then a whole segment of the scientific literature needs to be deposited in the wastebasket”. And we couldn’t have that, now, could we?
BA: We who? Me?
RM: I was being sarcastic, of course. Since we both have PhD’s in scientific psychology, what Bill was saying was directed at us; he was saying that all the work we did before understanding behavior in terms of PCT can be deposited in the wastebasket. I deposited mine there long ago (at least conceptually; I keep copies just to remind myself of the pre-PCT me); have you done the same with yours?Â
BestÂ
Rick
Richard S. MarkenÂ
"Perfection is achieved not when you have nothing more to add, but when you
have nothing left to take away.�
                --Antoine de Saint-Exupery
[From Bruce Abbott (2017.01.29.0840 EST)]
Rick Marken (2017.01.28.1750) –
Bruce Abbott (2017.01.28.1610 EST)
BA: The “standard PCT complaint� that Fred raised was that the computational plan-compute-execute approach is unable to deal effectively with the effects of unpredictable disturbances. Are you saying that this is not a standard PCT complaint against MCT?
BA: I noted that the sorts of disturbances Fred probably had in mind are actually handled by the vehicle’s low-level controllers – the ones that compare the vehicle’s position on the rroad to its reference position, compare the distance to the vehicle or other obstruction ahead to the reference minimum distance, and so on.
RM: The low level controllers themselves are very likely to be MCT controllers.
BA: Â Here is the relevant quote from the article:
The goal of Junior’s control system is to take the upcoming trajectory output by the planner and generate system inputs (throttle/braking and steering torque) in order to follow this trajectory. To achieve this end, we employ a mixture of a model predictive control (MPC) strategy, based upon well-known physically based vehicle models, along with feedforward proportional integral derivative (PID) control for the lower-level feedback control tasks such as applying torque to achieve a desired wheel angle.
Bruce