Thursday, 21 January 2016

A Quick Review and Analysis of Perceptual Control Theory

Perceptual Control Theory (PCT; Powers, 1973) is a theory that proposes behaviour is about the control of perception. We act so as to keep some perceived part of the world at some state, and it's by doing this to sensible variables that we come to exhibit functional behaviour. People have noted the seeming overlap between PCT and the ecological approach, and it's advocates (mainly Richard Marken and Warren Mansell) all talk about it in revolutionary terms that should also feel a bit familiar.

I first encounted it in the context of an interview with Richard Marken on a now defunct blog (pdf of the archived pagelink to page and scroll down to "Interview with Richard Marken"). Marken and I got into it a bit in the comments, as you will see! I was not impressed. However, Mansell & Marken (2015) have just published what they pitch as a clear exposition of what PCT actually is and how it works. I took the opportunity to read this and evaluate PCT as a 'grand theory of behaviour'.

My basic opinion has not changed. PCT is not wrong in most of it's basic claims, but it has no theory of information or how that information comes to be made or relate to the dynamics of the world. It's an unconstrained model fitting exercise, and it's central ideas simply don't serve as the kind of guide to discovery as a good theory should. Ecological psychology does a much more effective job of solving the relevant problems. 

Mansell & Marken first talk about stimulus-response behaviourism and cognitive theories as the two basic 'grand theories of behaviour', and point out (correctly) that they both treat behaviour as the output of a system. For example, when pressing a button, that press is the 'last step in a causal chain that starts in the nervous system and ends in the observed behaviour' (pg 426). Powers' insight was that the behaviour depends on both the actor and the environment. In order to produce a stable behaviour (e.g. a button press) the actor must constantly modulate the forces involved to cope with variations in the way the button responds. You get the same end from multiple means, and preserving an outcome in the face of variability and perturbation signals a process of control.


Figure 1. The basic sketch of Perceptual Control Theory
Figure 1 shows the idea. An organism is trying to control the perceived value p(t) of a variable y(t) so that it matches a target reference value r(t). You build your system so that it's activity works to control that perception and boom!, functional behaviour with respect to the controlled variable. 
If the system is properly designed so that the error drives the output in a way that pushes the controlled variable, and thus the perception of the controlled variable, towards the goal specification, then the goal result will be produced.
pg 427
The paradigm test case for PCT has, funnily enough, been the outfielder problem. Marken has some papers (e.g. Marken, 2005) modelling interception in terms of moving so as to control the perceived (optical) velocity of the ball and keep it at a reference value of zero. He can get this to work in a variety of interception tasks (animated demos here). I'm not sure this account can replace the existing OAC and LOT strategies; for example, Marken's model produces linear optical trajectories while controlling velocity and the evidence I think suggests people are actually moving to control the trajectory and not the velocity. 

An evaluation
This is the basic limitation of PCT; it has no theory of what perceptual information is and what kind of information is created in given tasks. This really shows up in the other thing Marken has modeled, bimanual coordinated rhythmic movement. Here he produces a model that manages to produce some of the effects seen in Meschner et al (2001), a Nature paper on the perceptual basis of coordination that completely fails to engage with any of the literature on that topic and that makes inaccurate claims about what that perceptual basis is. Marken's model works to control things like the relative angle between the flags in the experiment, their speed, and then finally the velocity of the hands. He produces some of the effects, but then at the end notes
The model is not yet able to capture the fact that participants producing antiphase rotations would often end up controlling for symmetry as they increased the speed of flag rotation. Apparently, control of the perception of symmetry is preferred to control of the perception of antiphase. The model does not yet explain this preference.
This is the switch from anti-phase to in-phase that is literally one of the signature features of coordinated rhythmic movements. His model can't account for it because he built it to control things that the actual system is not controlling (contrast this to the Bingham model). So he can fit some data, but not explain the behaviour (proper mechanistic explanations need real parts and operations and without a theory of what information actually is his model cannot explain anything). 

Further thoughts
While I don't think PCT is worth pursuing in and of itself, I do think this is a bit of a shame because I do think most of the basic analysis is right. We do organise our behaviour so as to build a task specific device ('a properly designed system') that, when controlled by information present in tasks, works in such a way as to complement the demands of that task. I don't think this requires internal representation of goal states (r(t)) but the essence of PCT is a closed loop control system, which is very ecological. I'm also finding the language of exactly what control means to be very useful in a paper I'm working on with Sabrina.

PCT is nowhere as influential as this article implies; based on their own references, there are just a handful of people doing PCT work and only two of them (the authors) are doing it properly. This makes some of their grand claims about the status and influence of PCT come across as just weird. My hunch is that PCT has not gone anywhere because it provides no guide to discovery. It is a sensible principle but with no theory or methods to shape how that principle is applied to specific behaviours. Right or wrong, the ecological approach provides this guide in spades and we have the empirical literature to prove it.

Summary
Perceptual Control Theory is interesting and it's basic insight (that behaviour is the result of closed loop control) is essentially correct. However, PCT is not the only theory to realise this. The ecological approach has this idea at it's core as well and has the added benefit of being able to guide discovery and empirical research. 

References
Mansell, W. & Marken, R. S. (2015). The origins and future of control theory in psychology. Review of General Psychology, 19(4), 425-430. Download ($$)

46 comments:

  1. PCT describes eleven levels of perception. These provide the specification for the information within the environment that is controlled. The levels are intensity, sensation, configuration (patterns), transitions, relationships, categories, events, sequences, programs, principles and system concepts. The list of levels is not exhaustive but Powers' best estimation. They specify the 'information array' in the environment that is to be controlled- known as the input quantity in PCT. It would be an interesting exercise to see how the perceptual levels map on to information in the environment that is specified in ecological psychology. But I see no contradiction. PCT is a macro theory and would not attempt to specify the kinematics of the environment in detail as these are likely to be specific to any particular environment, species and task. I am pleased that Andrew recognised a closed loop organisation is likely to be correct but I don't think he has considered its massive implications for theory, research and societal interventions. I would recommend reading Powers (1973) and running some of the demos at billpct.org.

    ReplyDelete
    Replies
    1. PCT is a macro theory and would not attempt to specify the kinematics of the environment in detail as these are likely to be specific to any particular environment, species and task.
      But this is my point. If you aren't specific about the information that is actually available to an organism, then you aren't modelling the actual mechanism working in a specific task.

      Take the coordination model: because you don't have people actually controlling relative phase perceived as relative direction then your model fails to produce the basic 0°/180° stability difference and you end up not describing what actually happens when someone engages in a coordinated rhythmic movement.

      So even though closed loop control is of course the deal (as the ecological approach also lays out, just with specific reference to what information closes the loop) PCT models fail to explain behaviour.

      Delete
  2. The information specific to a task is modelled on a task by task basis, not ignored. http://pctresources.com/CSGnet/Files/CSGnet_archive/CSGnet_attachments/Individual_files/Powers-1999-ModelControlledArmMovement.pdf. Or real robots are built circumventing the need to mathematically model the kinematics e.g. www.perceptualrobots.com. Or are you trying to say something else?

    ReplyDelete
    Replies
    1. That pdf link gives me a 404, sorry, so I can't see the point you want to make; can you summarise?

      Delete
  3. Let's add a few more points while Andrew is reading the paper I sent. To say that the ecological approach has always been 'closed loop' is not sufficient. The closed loop architecture in PCT is not any old closed loop. It has specific components, specified in the 1960 papers, that are organised in a specific manner. Each component has a specific function and the emergent purposeful behaviour is specific to PCT simulations (see billpct.org). My research on PCT is modest compared to that of the ethnologists Pellis & Bell, the sociologist Kent McClelland, the neuroscientist Henry Yin, and the organisational psychologist Jeffrey Vancouver. It is simply not a fair, accurate, kind or collegial position to dismiss their work. Finally I would hate to miss the opportunity to learn about how one might infer the kinds of environmental dynamics that lead to perceptual information presented to the senses that might be good candidates for controlled variables with a PCT hierarchy. I would value Andrew's insights for this. Yet Andrew needs to appreciate that a unified model of psychological functioning cannot specify the components of every situation. Moreover, he needs to realise that an ecological theory that focuses on specifying the environmental features of a behaviour cannot be a complete theory of behaviour. PCT provides the building blocks for the functioning of a nervous system that can utilise the perceptul information in the environment in a way that is complex, flexible and meets subordinate (e.g. Social, self-definitional goals). I wouldn't know where to start to use ecological
    psychology to treat a patient with psychosis, manage staff, predict group processes, model mental simulation, teach a class, bring up a child to name but a few. But PCT focuses on the core tenets and functional architecture I would require to do this - so the applications and evidence show. see PCTWeb.org

    ReplyDelete
  4. PS I meant 'superordinate goals' above!

    ReplyDelete
  5. A Slow Rebuttal to a Quick Review and Analysis of Perceptual Control Theory

    (Since the comments can only be 4096 characters this will have to be an Agony in Three Fits)

    Fit the First

    Andrew Wilson’s review of PCT was apparently motivated by the publication of an article by Warren Mansell and me (Mansell, W. & Marken, R. S. (2015). The origins and future of control theory in psychology. Review of General Psychology, 19(4),425-430). While I am flattered that Andrew was interested enough to read and comment on the article I think his review reflects some misunderstandings about PCT. These misunderstandings start in the first sentence of the review: “Perceptual Control Theory (PCT; Powers, 1973) is a theory that proposes behaviour is about the control of perception”. Actually, PCT is a theory that explains the fact that what we see as behavior is a process of control: keeping aspects of the world in pre-selected (reference) states, protected from disturbance (perturbation). The fact that behavior is control is explained in lay terms in my recently published book (with Tim Carey) entitled “Controlling People” (http://www.amazon.com/Controlling-People-Paradoxical-Nature-Being/dp/1922117641).

    Thus, it is the fact that PCT explains the phenomenon of control that distinguishes it from all other theories of behavior, including “ecological” theories, such as that described in the PowerPoint presentation “Ecological Information Makes Embodiment Possible” by Wilson and Golonka (https://cognitioninaction.files.wordpress.com/2015/10/2015-10-avant-talk.pptx ) Ecological theories and PCT are superficially similar because they are both “embodied” in the sense that they recognize that behavior depends on both the actor and the environment. Moreover, they recognize that consistent behavior is produced in an environment that contains perturbations that should prevent such consistency. The difference between these theories is that the ecological approach sees this consistency as a causal phenomenon, such as that seen in the behavior of “attractors” like a mass-spring system where a variable -- the position of the mass -- returns to the pre-perturbation state after removal of the perturbation – see slide 35 in the Wilson and Golonka PowerPoint presentation); PCT sees this consistency as a control phenomenon such as that seen in the behavior of a control system like a thermostat where a variable -- the temperature of the room – remains in a reference (non-perturbed) state protected from the ever-present variations in perturbations, such as variations in the number of people in the room.
    You go on to say: “PCT is not wrong in most of its basic claims, but it has no theory of information or how that information comes to be made or relate to the dynamics of the world”. But PCT doesn’t have these things because it doesn’t need them. Apparently you think PCT should have these things – a theory of information or a theory of how that information comes to be made and of how it relates to the dynamics of the world – because your theory has them. I think theories should be evaluated in terms of their ability to account for the data, not in terms of whether they are like your own theory.

    Next you say: “[PCT] is an unconstrained model fitting exercise, and its central ideas simply don't serve as the kind of guide to discovery as a good theory should.” I don’t understand why you think PCT is “unconstrained”. I think it’s rather severely constrained by the equations that defined the model. And PCT provides a very clear guide to discovery, though it’s focused on discovering the variables that organisms control – controlled variables. A rather nice Guide to Discovery in the context of PCT can be found starting on p. 167 in a collection of papers by W. T. Powers entitled “Living Control Systems” (available at
    http://www.amazon.com/Living-Control-Systems-Selected-Papers/dp/0962415405/ref=sr_1_4?s=books&ie=UTF8&qid=1453656589&sr=1-4&keywords=living+control+systems).

    ReplyDelete
    Replies
    1. The difference between these theories is that the ecological approach sees this consistency as a causal phenomenon, such as that seen in the behavior of “attractors” like a mass-spring system where a variable -- the position of the mass -- returns to the pre-perturbation state after removal of the perturbation
      First, ecological psychology uses dynamical systems theory as a tool but not as an explanatory theory. Attractors aren't things that have causal power; in the model, the system gets to take advantage of some free control offered by the mass-spring dynamic but the 'attractor' that represents control is actually implemented using information.

      Apparently you think PCT should have these things – a theory of information or a theory of how that information comes to be made and of how it relates to the dynamics of the world – because your theory has them. I think theories should be evaluated in terms of their ability to account for the data, not in terms of whether they are like your own theory.
      No, I think you need them because you need them! Without them, PCT is just a model fitting exercise and won't lead to genuine mechanistic explanations (same with standard cog sci, etc). If you aren't trading in real parts and operations in your models, then your models can't explain anything, only describe. We think we should aim higher and that we can thanks to ecological information.

      Hence 'unconstrained': maybe it would be clearer if I said you aren't constrained by the right things. Parsimony, data fitting, etc, aren't the right criteria: the constraint on whether a variable eneters a model should be 'does it represent a real part or operation of the mechanism under investigation?' See Bechtel and Abrahamsen for more on this.

      Delete
    2. AW: Attractors aren't things that have causal power; in the model, the system gets to take advantage of some free control offered by the mass-spring dynamic but the 'attractor' that represents control is actually implemented using information.

      RM: How does the system “take advantage” of the “free control” offered by the mass-spring dynamic? What is the mechanism of this “taking advantage” process? And what “free control” is offered by the mass-spring dynamic? Equilibrium systems, like a mass on a spring or a pendulum, don’t control at all; a perturbation, such as a displacement of the position of the mass, is fully effective (per Hooke’s law); the mass returns to the equilibrium position only after the perturbation is removed.
      AW: I think you need them [a theory of information or a theory of how that information comes to be made and of how it relates to the dynamics of the world] because you need them! Without them, PCT is just a model fitting exercise and won't lead to genuine mechanistic explanations (same with standard cog sci, etc). If you aren't trading in real parts and operations in your models, then your models can't explain anything, only describe.

      RM: I heartily agree with your last sentence. But PCT does “trade in real parts” and it describes the mechanism that is implemented by these parts. Take a look at the PCT model of the Mechsner coordination task at http://www.mindreadings.com/ControlDemo/Coordination.html. The component s of this model correspond to “real parts” of the organism and the environment: the moving flags (in the environment) are sensed by the eye; neural networks in the optic path compute perception of the relative angles of the flags; the perception is an afferent neural signal; this perception is compared (via synaptic connections) to a reference signal, which is an efferent neural signal. The resulting error signal is also an efferent neural signal that drives the muscles that result in movement of the hands (in the environment) that cause the flags to move. Every component of this model is consistent with what we know about the nervous system and the physics of the environment. The mechanism carried out by these “real parts” is specified by the equations that describe the functional relationships between the parts – functional relationships that are theoretical on the organism side and known physical relationships on the environment side.

      So I’m surprised by your claim that PCT is a non-mechanistic model that doesn’t trade in “real parts”. Indeed, it is my impression that “non-mechanistic” and “doesn’t trade in real parts” is a better description of your model than of PCT. I have looked at the papers by Bechtel and Abrahamsen that you recommended but frankly I couldn’t see the “real parts” to which those models correspond or even the mechanism that makes the model work. I think the problem is that I have seen no functional diagram of your model – a diagram like Figure 1 in our paper that was reprinted in your Review. What would really help me understand your criticism of the PCT model would be a diagram, equivalent to the one on my “Bimanual Coordination” page (http://www.mindreadings.com/ControlDemo/Coordination.html), showing how your model explains the behavior observed in the Mechsner task described on that page. The diagram should make it easier for me to see how your model works (the mechanism) and how it corresponds to the real parts of the organism performing the task and of the environment in which the task occurs.

      Delete
    3. On real parts and operations:
      neural networks in the optic path compute perception of the relative angles of the flags
      So this is not a real part. It's a real thing, obviously, the flags have these angles. But it's not part of the mechanism of a coordinated rhythmic movement.

      Bingham's model uses relative direction as the information for relative phase, and relative speed as the noise term. Our joint empirical work has confirmed all of this.

      So when I saw real parts and operations, I don't just mean 'things that real exist' I mean 'things that actually take part in the mechanism'. Sorry if that wasn't clear!

      Delete
    4. How does the system “take advantage” of the “free control” offered by the mass-spring dynamic? What is the mechanism of this “taking advantage” process? And what “free control” is offered by the mass-spring dynamic? Equilibrium systems, like a mass on a spring or a pendulum, don’t control at all; a perturbation, such as a displacement of the position of the mass, is fully effective (per Hooke’s law); the mass returns to the equilibrium position only after the perturbation is removed.
      Fair points, I was being imprecise with the term control.

      The 'free' bit you get from the composition and organisation of limb dynamics (eg the assembly of a limb into a device best modelled as a damped mass-spring) is a reduction in the control requirements. If your limb is a damped mass spring, you can move that spring by shifting the equilibrium point (cf Feldman) and not have to actively control the limb at all points throughout the trajectory as it moves from point A to point B.

      Assembling the action system into a task appropriate low dimensional 'task specific device' or 'synergy' is a critical part of the process of reducing the degrees of freedom currently operating to a controllable number.

      Delete
    5. I guess I'm not going to get what I want -- a diagram (or, better yet, a computer program) showing how your model produces the behavior observed in the Mechsner experiment, which you claim it does and which I believe it can't. I don't believe it can produce _any_ of the results observed in the Mechsner experiment, including the reversal from antiphase to symmetrical flag movement at higher speed. But never mind the reversal that you say is so significant (how often does this happen in real life? do pianists, for example, playing two rapid, simultaneous contrapuntal voices regularly lapse into playing the voices in symmetry?), unless you can show me how your model produces just the symmetrical and antiphase motions of two flags using hand movements that are neither symmetrical nor antiphase (that is, unless you can show me that your model can do what the PCT model does in my demo (http://www.mindreadings.com/ControlDemo/Coordination.html)I will continue to believe that your "model" is nothing more than a curve fitting exercise -- the very thing you accuse PCT of being.

      Delete
    6. The model is described here. You can code it up yourself. Geoff has it implemented in Simulink, in Matlab, but anything like that or Labview can handle it.

      That paper details how the model does indeed produce all the relevant phenomena.

      The switch from 180° to 0° really is literally a hallmark of coordination dynamics. It's caused by perceiving relative phase in terms of relative direction, so anyone who needs to produce a wide range of relative phases shifts information variables (to, for example, relative position, which doesn't lead to the switch). These aren't simply things I am claiming. They are keys parts of the task dynamic with endless data about it.

      You can believe what you like, but the model and it's abilities have been tested, described and published in a variety of papers.

      Delete
    7. I can't tell from that "model" described in that paper how to make it behave like the subjects in the Mechsner experiment. I'm pretty sure it can't. So while your model may be able to account of the transition from anti-phase to symmetry at high frequencies mine can account for the ability to to produce symmetrical and anti-phase flag movements using bimanual movements whose phase relationship is continuously changing. So the score is currently 1 to 1. I think we have to go into overtime;-)

      Delete
    8. Read Wilson et al 2005a for experiments based on the model about getting various phases in things like flags (dots on screens for me) from movements at different phases. I do a much more comprehensive job than Meschner et al and I have an explanation, based in the model.

      Delete
    9. Also I wouldn't fixate on that paper. It's one, not very good paper in a wide literature detailing the basic phenomena defining coordinated rhythmic movement. Geoff's model engages with that literature; you should too.

      Delete
  6. Fit the Second

    You give a pretty good recap of the PCT model, as described in Figure 1 of our paper. But for some reason you seem to scoff at the ability of the model to handle the “outfielder problem”. You say “I'm not sure this account can replace the existing OAC and LOT strategies; for example, Marken's model produces linear optical trajectories while controlling velocity and the evidence I think suggests people are actually moving to control not trajectory and not the velocity”. Actually, the evidence strongly suggests that people control, not trajectory but vertical optical velocity and lateral optical displacement. This is described in a couple of recent papers describing a PCT model of object interception (which includes intercepting fly balls). In one case, people intercepted toy helicopters that have very irregular trajectories and trace out quite non-linear optical trajectories (https://www.dropbox.com/s/s2p2bheqv9f5el3/Chasin%27Choppers.pdf?dl=0). Another describes interception of objects shown to oneself: (https://www.dropbox.com/s/kng5it0psk0t10v/SelfCatch.pdf?dl=0). In both cases the PCT model (called COV for Control of Optical Velocity) did better than OAC and LOT.

    By the way, I did the modeling using data collected by Dennis Shaffer, who was one of the developers of LOT and went on to develop a “segmented” version of LOT (called SLOT) in an effort to bail out LOT when it turned out that the optical trajectories objects moving in highly variable actual trajectories are not linear. Actually LOT and SLOT are not really models in the PCT sense since they are merely exercises in curve fitting. Talk about “unconstrained”. It turns out that PCT does a pretty darn good job of accounting for all kinds of object interception behavior. And it does it by, not by “unconstrained” curve fitting but, rather, by controlling perceptions in the same environment as that in which the actual subjects had done this controlling. I myself was quite surprised at how well the simple control model, controlling two different perceptual variables – optical velocity and lateral displacement – accounted for behavior in many different object interception situations.

    Finally, you again claim that the basic limitation of PCT is that “it has no theory of what perceptual information is and what kind of information is created in given tasks”. You say that this limitation “really shows up in the other thing Marken has modeled, bimanual coordinated rhythmic movement. Here he produces a model that manages to produce some of the effects seen in Meschner et al (2001)…Marken's model works to control things like the relative angle between the flags in the experiment, their speed, and then finally the velocity of the hands”. And that’s all it was designed to do. I chose to model the behavior in this experiment because the coordination was perceptual, not manual. When you play the simulation (at http://www.mindreadings.com/ControlDemo/Coordination.html) you will see that while the movement of the two flags is coordinated the movements of the two hands are not. So this is not a case of bimanual coordination but of bi-perceptual coordination. The control model shows that coordination is organized around the control of input (perception), not output (hand movements, in this case).

    You say: “[The PCT model] produces some of the effects [observed in the Mechsner experiment]…” In fact, it produced the effect that was most important to me; the ability to produce a coordinated result (symmetrical or antiphase flag movement) using uncoordinated means (uncorrelated hand movement). I would be interested in seeing whether your “ecological” model could do this. I am pretty sure that it can’t, unless it is a control of perception (PCT) model.

    ReplyDelete
    Replies
    1. Effective feedback but not an effective link! Try this

      Delete
    2. Just to clarify: those papers are me replying to
      I would be interested in seeing whether your “ecological” model could do this. I am pretty sure that it can’t
      We can account for it, I have run similar experiments (eg Wilson et al 2005a, b) and I have other studies detailing how and why it works in terms of the model.

      Delete
  7. Fit the Third

    Although my model successfully accounted for nearly all of the results of the Mechsner study – with no indication that your model could do the same – you chide me for the inability of my model “to capture the fact that participants producing antiphase rotations would often end up controlling for symmetry as they increased the speed of flag rotation”. You say that “This is the switch from anti-phase to in-phase that is literally one of the signature features of coordinated rhythmic movements”. This may indeed be a “signature feature” – it is very interesting -- but, as you note, it is only one such feature. It seems to me that the coordinated result is itself rather signature. It would have been nice if the switch from antiphase to symmetry “fell out” of the model. But it didn’t. Perhaps your model can help here. It would be great if you could show me how your model accounts for the results of the Mechsner experiment, including the shift from antiphase to symmetry at high rotation speed. You say that my model “can't account for [the shift] because he built it to control things that the actual system is not controlling (contrast this to the Bingham model)” implying that the Bingham model can account for all the data, including the shift. I would really like to see how that works.

    I appreciate your final efforts to say something nice about PCT, but I’m afraid that you are complimenting PCT for the wrong reasons. You say: “While I don't think PCT is worth pursuing in and of itself, I do think this is a bit of a shame because I do think most of the basic analysis is right. We do organise our behaviour so as to build a task specific device ('a properly designed system') that, when controlled by information present in tasks, works in such a way as to complement the demands of that task.” But PCT doesn’t say this at all. There is nothing in PCT about behavior or task specific devices or properly designed systems being “controlled by information”. In PCT it is perception (or, if you like, “information”) that is controlled. In PCT, behavior is the control OF information; behavior is not controlled BY information. This fact is demonstrated in spades in the Mechsner experiment; there is simply no information in the movement (or relative movement) of the flags that can control the hand movements so that the flags move in symmetry (or antiphase). The hand movements control information about the relative movement of the flags, not vice versa.

    You are certainly correct when you say that “PCT is nowhere as influential as this article implies…”. If the article implies that PCT has been influential then this is very misleading. What we were trying to say is that PCT has influenced a lot of work in the field of “self-regulation” but, unfortunately, none of that work is correctly informed by PCT. It uses the terminology of PCT but not the basic understanding that the behavior of living systems is control, that it is organized around the control of (not by) perceptual variables and that the main goal of research aimed at understanding behavior of organisms is to discover the variables they control. This is a paradigm shift up with which few psychologists have been willing (or able) to put.

    You conclude by saying “My hunch is that PCT has not gone anywhere because it provides no guide to discovery. It is a sensible principle but with no theory or methods to shape how that principle is applied to specific behaviours.” Again, I call your attention to the PCT Guide to Discovery starting on p. 167 of “Living Control Systems”. I also have a little collection of papers that are all relevant to question of how to do research in the context of understanding living systems as control (purposeful) systems. The book is called “Doing Research on Purpose” and you can be the 3rd person in the world to own a copy (my wife and my mother are the other two) by going to http://www.amazon.com/s/ref=nb_sb_noss_2?url=search-alias%3Dstripbooks&field-keywords=doing+research+on+purpose

    Best regards

    Rick

    ReplyDelete
    Replies
    1. Geoff's model, which does indeed do all the things you mention. I reviewed it in some detail here

      Delete
  8. Hi Andrew, I am waiting for a response to my last comment and Rick's comments. Particularly one that (a) addresses the points in turn; (b) doesn't involve links to papers that are not directly related to the issues of whether PCT is an accurate macro theory of behaviour which is what I understand to be the point of your blog entry; (c) includes components specifications of an alternative theory that either work as a simulation or as a robotic device; (d) are somewhat wider than the issue of bimanual coordination that happens to be your research domain. All the best, Warren

    ReplyDelete
  9. So while I am intrigued by the idea of behaviour as the side effect of controlling perception, I am left with a question:

    If I am not trying to control actions, which variables do I act so as to control at any given time, and why?

    ReplyDelete
  10. I am pleased you are intrigued! 'Control' is defined as keeping a variable at a preselected value despite disturbances to that variable. Therefore you control only your perception at any one point. For example you may have a perception of your current movement velocity extracted from the 'visual flow' as you move forward, that is continuously compare with your reference speed. To do this despite disturbances (friction, wind resistance, body inertia, other muscles, gravity) you need to vary your behaviour dynamically - for example putting in more force when going up a hill. Thus, muscle forces are 'commanded' but not controlled. This one level example is simplified. It should be possible to break down any coordinated perception into hierarchical levels in which each level controls its input by sending reference signals to the level below. Only the lowest layer interfaces with the environment. Diagrams help of course! See Powers (1973, 2008).

    ReplyDelete
  11. Now the why question. The CV of optical velocity is just one example, selected for the superordinate goal at hand that happens to match the current situation. So if there is a more convenient or reliable CV, this will be used. For example there is less need to use optical velocity when one has s speedometer in a car. In real life probably both are used. The CV and its current reference value is selected also by the output of a Superordinate system which in turn is selected by a higher level system, etc. one might control one's perceived speed to 'be a law abiding citizen' or 'to be approved by my passenger'. Why might one wan to perceive a high level of approval from a passenger? Maybe it's your wife and you want to have a good relationship. Why want to experience a good relationship? Here we come to near the top of the hierarchy. Often people say 'it just feels right or feels good'. Here we hit the intrinsic reference states for living things. PCT described how we develop our multilayered perceptual hierarchy in early life to fulfil our intrinsic needs. During adulthood we rely on these less frequently. Frans Plooij developed the popular patenting book Wonder Weeks based on decades of observation and interpreted the stages of development as layers of the PCT hierarchy. Is that the answer you are looking for? Note that the environment is the source of the information that is perceived at all of these levels (although mental simulation can draw on stored envtal information). So the environment is critical, as part of the loop, but it is the internal specification of this environmental information that is directly controlled.

    ReplyDelete
    Replies
    1. Each of those higher order systems is essentially just another negative feedback loop as per the figure, right? So it's control loops all the way 'up', controlling progressively higher and higher order variables.

      So what sort of thing is a perceived variable for the loop controlling 'be a law abiding citizen'?

      Delete
  12. Additional notes. 1. I have skipped levels (11 in total) for ease of explanation. 2. The CVs for higher level systems are schematic not propositional even though I have represented them in words here for ease. 3. I have explained why CVs are set but not how. This would require some discussion of the nature of distributed memory in PCT and of its learning alogirithm known as reorganisation. See pctweb.org. Or I could continue?

    ReplyDelete
    Replies
    1. Oh, didn't see this;
      The CVs for higher level systems are schematic not propositional even though I have represented them in words here for ease
      The question stands I guess; what sort of thing is a controlled variable for those higher order loops?

      Delete
  13. Hi Andrew, thank you again for your curious questioning. First, I have a diagram that may help get across the nature of the levels and there continuity: http://www.pctweb.org/LevelsofHPCT.pdf
    Second, to answer verbally. Every level in the hierarchy is an abstraction of incoming sensory input from the environment. The higher levels are no different. In the psychotherapeutic literature this schematic higher level perception of the self in the present moment is often called the 'felt sense'. See for example http://www.courtenay-young.co.uk/courtenay/articles/Encouraging_the_Felt_Sense_of_Self.pdf
    Of course I would not rule out that language could have a role in the perception of higher level CVs but it may in fact have a different one, as proposed by Powers - a way of using symbolic perceptions to 'stand in' for felt senses thereby allowing communication about these perceptions to be easier and less emotionally sensitive (I.e. demanding of current attempts to control them; e.g. We can talk about 'terror' without needing to experience it).
    Personally as a therapist, I often use mental imagery to help people 'crystallise' their desired perceptions of themselves in the world and this acts as a template for their value-driven action.
    Does the diagram and this interpretation help?

    ReplyDelete
    Replies
    1. This all sounds a little cognitive to me, in that it very quickly gets away from actual parts and processes (optical velocity is an actual thing you can find and do experiments on) and into things that are more like functional capacities ('felt sense' sounds perceptual only in a fairly analogy kind of way; what sort of information variables support it?) Is that fair? That's ok, it's just not what we're doing here and I'm just trying to map the lines out a bit.

      Delete
  14. Well you asked what was at the top, and most humans I know are driven ultimately by those kind of abstract values rather than by an arbitrary choice of their preferred speed of movement! The same won't apply to simple/very young animals of course! Also note that there are nine 'concrete' levels below the top two 'abstract' ones. See the hierarchy diagram linked in my last post. These are explained very clearly in the book Making Sense of Behavior: The Meaning of Control (Powers, 1998). But the 1960 and 1973 publications include earlier versions. It's probably too much to go through each in detail here.

    ReplyDelete
    Replies
    1. most humans I know are driven ultimately by those kind of abstract values rather than by an arbitrary choice of their preferred speed of movement!
      1. The set values people choose for informational control like this aren't arbitrary. They have to be tied to the task dynamics or else the behaviour won't be functional.

      2. It's called 'Perceptual Control Theory' and the basic architecture involves working to maintain some perceived value at some reference state. So what's being perceived and therefore controlled at those higher levels?

      Also, how do you deal with feedback delays?

      Delete
  15. I’ll take the liberty of replying to this. I’ll use initials to show who is saying what.

    AW: 1. The set values people choose for informational control like this aren't arbitrary. They have to be tied to the task dynamics or else the behaviour won't be functional.

    RM: I presume “set values” refers to what are called “reference values” or actually “reference signals” in PCT. Indeed, they are not chosen arbitrarily; their value is set by higher level systems as the means of controlling those higher level systems’ perceptions. Some time ago I developed a spreadsheet model of a hierarchy of perceptual control systems to illustrate how this works. The model is described in Marken, R. S. (1990) Spreadsheet Analysis of a Hierarchical Control System Model of Behavior, Behavior Research Methods, Instruments, & Computers, 22, 349 - 359. It was developed in Lotus 1*2*3 but I have an Excel version online at http://www.mindreadings.com/ControlDemo/hierarchy.xlsx.

    RM: And it is not true that these reference values (signals) have to be tied to task dynamics. What you call “task dynamics” are, in PCT, the physics of the feedback function that connects control system output (the neural signals going into the muscles) and sensory input (on which the controlled perception is based). What does have to be appropriately “tied” to these task dynamics are the parameters of the control system that influence the dynamics of its behavior. These are the gain, time constant (integration factor) and transport lag of the system. From a PCT perspective, learning to perform a task is, to a large extent, learning the appropriate settings for these parameters.
    AW: 2. It's called 'Perceptual Control Theory' and the basic architecture involves working to maintain some perceived value at some reference state. So what's being perceived and therefore controlled at those higher levels?

    RM: That’s really what PCT research would be about; figuring out what kinds of perceptions are controlled at different levels (or in different location – it might not be a hierarchy, though so far the evidence strongly suggests that it is). Powers has hypothesized that there are 11 levels of control, each level controlling a different type of perceptual variable. The lowest level type of perception is an intensity variable, such as the degree of tension in a muscle fiber. The next level is a sensation variable – a combination of intensities—such as the sensation of tension in a muscle. The next level is a configuration variable – a function of sensations – such as the perception of the angle of the arm. The next level is a transition variable – the change in a configuration (or sensation or intensity) over time; it’s the perception of rate of change; a derivative. Above transitions are relationships, events, sequences, programs, principles, system concepts.

    RM: Again, these ideas about the types of perceptual variables people control and the hierarchical relationship between them are hypotheses around which research on PCT can be organized. And we have done some research based on these hypotheses. A nice example is in Marken, R. S., Khatib, Z. and Mansell, W. (2013) Motor Control as the Control of Perception, Perceptual and Motor Skills, 117, 236-247. That study provides evidence that people control configuration, transition and sequences and that these variables are hierarchically organized in that order. I have also done some unpublished work showing that people can control program type percpetoins. I have a demo on the net that shows what it means to say that people control these different types of perceptual variable: http://www.mindreadings.com/ControlDemo/Hierarchy.htm


    Rick

    ReplyDelete
  16. Sorry, due to word length constraints I had to do this in two posts. Here's the answer to Andrew's final question:

    AM: Also, how do you deal with feedback delays?

    RM: There are two kinds of “delay” in a feedback loop. One is the transport lag introduced by the nervous system itself; the other is the slowing that can be introduced by the physics of the environment, such as the lag between moving the tiller (output) of a ship and the arrival of the ship at the intended direction (input). Both of these sources of “feedback delay” are handled by appropriate setting of a control system’s gain and time constant. The feedback delay introduced by the transport lag of the nervous system is never long enough to create any stability problems but the slowing introduced by the physics of the environment definitely can (if you’ve ever tried to steer a large ship or, better, dock the shuttle with the international space station, which I tried in simulation with disastrous results).

    RM: Hope this helps.

    Best

    ReplyDelete
  17. I empathise with Andrew because it is very tempting to propose that that the environment instructs the internal organisation of behaviour. Surely there is something about the nature of the motion of a baseball in the air that determines what controlled variable is used by the fielder to catch it? Yet, Rick's research shows that altering the dynamics of the object to be intercepted has a minimal effect on the architecture required to catch it effectively.
    The argument is analogous to the difference between Lamarkism and Natural Selection. According to Darwin, the wonderful match between the morphology of organisms and their functions in the environment is not a consequence of the environment instructing the process of evolution. Rather, the environment sets out a problem space within which the organism internally generated a range of its own possible solutions. To put it bluntly, the fact that trees on the Savannah are tall does not instruct a giraffe to have a long neck. Rather, this physical challenge in the environment sets out a challenge for the trial and error changes in generations of pre-giraffes to solve. Likewise, PCT recognises that within the lifespan of an organism, the physics of the environment cannot 'come in' and rewrite the organism. That does not deny that an external theorist cannot see the obvious link between the environmental
    Constraints and the morphology. Rather the rewiring happens Internally and spontaneously and those modifications that improve control are maintained. In PCT this is known as reorganisation and is analogous to natural selection over much shorter timescales (within a developmental period of an organism). Thus, there is an important philosophical issue we are wrestling with. Yet it is also empirically testable.

    ReplyDelete
  18. PS It is uncontentious that the immune system works in this way, so why not the nervous system? (Cziko, 1997)

    ReplyDelete
  19. Forgive me for resuscitating this comments section, but would it be correct to say that another difference between ecological theories and perceptual control theories is that the former deny that cognition entails internal representations (in the cognitivist sense) while the latter assumes some form of mental representation to account for feedback and control? Thanks in advance for the clarification.

    ReplyDelete
    Replies
    1. The big difference is that ecological psychology has a theory about the form of the information variables, while PCT just inserts something useful. This is a non-trivial difference, I think.

      Delete
    2. Do you mean to say that PCT positing internal representations is a trivial difference? But I thought the MAIN difference between extended/enactive accounts and other approaches (Bayesian included) is the jettisoning of representational theories?

      Delete
    3. I'm actually not sure where PCT stands on representations. I think they are probably lurking in there, though, so this would remain a difference.

      Delete
  20. I think the big difference between PCT and ecological theories is that they are about different things. PCT is an explanation of the controlling done by living systems; ecological theories aren't. PCT was developed to explain the fact that organisms maintain variable aspects of their environment in reference states, protected from variable disturbances. Ecological theories were developed to explain something else. So any apparent similarities or differences between the theories are completely coincidental.

    ReplyDelete
  21. Richard, would you say PCT requires in principle a representational model of nervous systems to account for an organism's ability to control the perceived value of a variable so that it matches a target reference value? For example, does it necessitate the received view of visual perception, such that perceptual inputs are filtered through brain regions in a top-down manner, with stored memory acting upon sense data to continuously update a parsed reconstruction of the environment to inform real-time behaviours. Or is PCT epistemically neutral in this regard, and theoretically amenable to direct realist models of perception-action? Thank you both for your responses.

    ReplyDelete
    Replies
    1. Hi Tim. The PCT model of perception is neither "representational" nor "realist" (as I understand those terms). The name we usually use for the PCT model of perception is "constructivist". In the PCT model, perceptual functions act to construct, from the sensory effects of the physical environment, the perceptual variables that the organism controls. So perceptual variables in PCT can be thought of as functions of physical variables -- variables like mass, force, electromagnetic energy, etc. that are part of our models of the physical world. In PCT, perceptual functions in organisms work like those in artificial control systems. An example is the perceptual function in an artificial humidity control system; this system constructs a perception of humidity from sensed measures of temperature (rate of molecular motion) and water vapor content in the air. This is an interesting perception because it does not correspond to an entity out in the environment (as far as our models of the physical world are concerned); but it's a perceptual variable that matters to people so we have designed artificial systems that can perceive and control this constructed variable, keeping it at a level that we perceive as comfortable.

      By the way, the PCT model of perception is rather nicely laid out in the basic book on PCT, William T. Powers "Behavior: The Control of Perception".

      Best

      Rick

      Delete
    2. So to answer the question, PCT is representational and quite cognitive, making this another point of departure from the ecological approach.

      Delete