Behavioral illusion, based on Powers (1978)

It comes down to the fact that PCT research cannot ignore the side effects of control. If there are regularities in behavior that are not the controlled variabe itself, a model must reproduce those regularities, and by definition, they are side effects.

The side effects will fall out of a model that has the correct controlled variable and the correct parameters, and is in an identical situation as the human subject.

If the regularities don’t fall out, well, then the model is not correct, regardless of why the model was developed, or what was the aim during development.

There is more in the archives on the Atkeson and Hollerbach paper. Look at BP (920330.0800), for a list of needed changes that would make the arm model trajectories conform to those observed in human behavior, as reported in A&H. Bill talks about a adding a possible path control level, and about adding a remapping between visual and kinestethic space, and about problems with raping movements.

RM: Thanks for that! I just scanned it quickly but i see that Bill was suggesting the addition of another, higher level perception to control. The recent research we have done on this – that will be part of the paper I’m writing that is inspired by this wonderful dialog – shows that the Little man model, as it existed when Bill built it, will produce invariant trajectory profiles that match the A-H profiles perfectly, which no need for a new level. But this post fro Bill is certainly worth thinking about.

Thanks again.

Best

Rick

Another thing to consider is that there are many different configurations that will produce a bell-shaped velocity profile (second-order damped dynamic systems). If you google for the VITE model, DIRECT model, NOCH model, and a whole bunch of other ones, they all reproduce the bell-shaped velocity profiles, as a sort of a basic test that any model needs to pass. The strength of the Little Man model is in the simplicity of control structures, and thus the higher biological plausibility (I think). They are all wonderful models, but maybe a bit too complex.

The weakness of the Little Man is that there is no “trajectory control”. I put it in quotes because most models have explicit control of trajectory. built in, and can reproduce any given velocity profile or geometric shape given. The little man, in contrast, can only follow accurately very slow reference trajectories, and this suggests, as Bill was saying, that there is another level of control missing from the little man, that does not care about trajectory as a single controlled variable, but about the path and some kind of speed independently, as two different controlled variables.

(I’m also working on a paper related to the little man demo)

Hi Adam

AM: Another thing to consider is that there are many different configurations that will produce a bell-shaped velocity profile (second-order damped dynamic systems). If you google for the VITE model, DIRECT model, NOCH model, and a whole bunch of other ones, they all reproduce the bell-shaped velocity profiles, as a sort of a basic test that any model needs to pass.

RM: I found VITE and DIRECT. They look like output-generation models to me. If they are, then it would be easy to show the superiority of the PCT model; just apply gentle and slow disturbances to the movements made by all the models (and the human subjects) doing the A-H task. Those disturbances should have little or no effect on the shape of the velocity profiles produced by the humans and the PCT model but should be fully effective on the shape of the profiles produced by VITE, DIRECT and all the other models. Remember, PCT is a model of controlled input, not of caused output.

AM: The strength of the Little Man model is in the simplicity of control structures, and thus the higher biological plausibility (I think). They are all wonderful models, but maybe a bit too complex.

RM: My guess is that the strength of the Little Man model is that it is a control model not a caused output model.

AM: The weakness of the Little Man is that there is no “trajectory control”.

RM:I’m not sure that’s true. Our work with the Little Man suggests that it accounts for the invariant velocity profiles without the addition of trajectory control. But that’s just for the task A-H had their subjects doing. Those profiles are just a side effect of controlling various perceptions so it’s not clear that those profiles would even exist in that form if subjects were put in a situation where they have to control those perceptions in a way that would result in a different shape of velocity profile as a side effect.

AM: I put it in quotes because most models have explicit control of trajectory. built in,

RM: I doubt it. The don’t have control of trajectory built in unless they are actually control models, controlling a perception of trajectory. My guess is that built into these models is a trajectory waveform that causes (generates) the output trajectory. Remember, cause and control are not synonyms!

AM: (I’m also working on a paper related to the little man demo)

RM: Great. I look forward to seeing it.

Best

Rick

RM: I found VITE and DIRECT. They look like output-generation models to me. If they are, then it would be easy to show the superiority of the PCT model; just apply gentle and slow disturbances to the movements made by all the models (and the human subjects) doing the A-H task. Those disturbances should have little or no effect on the shape of the velocity profiles produced by the humans and the PCT model but should be fully effective on the shape of the profiles produced by VITE, DIRECT and all the other models. Remember, PCT is a model of controlled input , not of caused output .

It is exactly opposite. It velocity profile is the controlled variable in humans, then disturbances will not have an effect. If the velocity profile is not controlled, then disturbances will have an effect on the velocity profile. If temperature in the room is controlled, then opening the window will not have much of an effect, because the AC will work stronger, and vice-versa.

VITE, DIRECT, NOCH, REACH, etc. are interesting and wonderful models to me, in the sense that they show different ways of producing hand trajectories, they look into biological structures, human behavior, etc. The have some strange feedback structures, I would not call them “caused output”, but they do use inverse kinematics and dynamics which might be overly complicated. There is also learning and adaptation, etc, which is very interesting if someone is into those kinds of things.

Some of them do have explicit trajectory sensing and control, rejecting disturbances to the trajectory, etc.

All of those researchers are trying to find good models of human hand and arm control, and they have varying degrees of success, just like the people in the PCT community. Some things they might be doing good, some things maybe less good, but so am I, and so are we.

AM: The weakness of the Little Man is that there is no “trajectory control”.
RM:I’m not sure that’s true. Our work with the Little Man suggests that it accounts for the invariant velocity profiles without the addition of trajectory control.

It sure does not require trajectory control to make nice bell-shaped velocity profiles, but neither does a damped mass on a spring. Lots of systems show bell-shaped velocity profiles. It is not too impressive or convincing is what I’m saying.

What is missing in the broader sense of growing the model toward biological realism, is a way of producing different trajectories seen in humans, such as drawing letters, or producing rhythmic movement, etc. It is not clear what are controlled variables in those behaviors, Bill hints it might be paths, for example, but he never programmed that level into any of the models.