Control Efficiency

[From Rick Marken (2006.07.06.1040)]

Rick Marken (2006.07.05.1845)--

I think a quantitative analysis of all this could be done using the equations for a quasi-static analysis of control described in Powers' 1978 _Psychological Review_ paper. But they could be used only if we have a quantitative definition of efficiency. Maybe I'll give it a shot (using my own definition of efficiency) sometime before I leave for China. It is kind of an interesting question: does the efficiency of control (I would say in terms of how much energy it uses to produce a particular level of control) vary depending on the parameters of control?

OK, let me try this. I'll base this analysis on Bill's difference equation approach to determining the relationship between loop gain and speed in a control system. The relevant difference equation is the one that computes change in output per unit change in input. The basic equation is

(1) qo(t+1) = qo(t) + K[F(qi(t)-qi*)-qo(t)]

This is equation 15 in the Psych Review article. qo (t) and qo(t+1) is the output variable value (think of it as the horizontal position of a mouse) at time t and t+1 respectively. K is a slowing factor (between 0 and 1) indicating the proportion of the new output that should be add to the current output to produce the output at the next instant, t+1. The larger the value of K the faster the output changes over time. F is a linear multiplier that corresponds to the "system amplification function". The larger F, the more "sensitive" the system is to error, (qi(t)-qi*), so the more output the system produces per unit error. qi and qi* are the input variable and reference for the input variable, respectively. In order to solve the close loop version of this equation, we have to take into account the fact that qi(t) = Gqo(t) + Hqd, where G is a linear multiplier that corresponds to the feedback function and H is the function that relates the disturbance variable, qd, to the input variable, qi(t).

Making all the proper substitutions and simplifications I get the following equation for that describes a change in output during one time unit:

(2) qo(t+1)-qo(t) = K(FG-1) qo(t) + KF (Hqd - qi*)

Now here's where the subjective interpretation begins. I take the multiplier of first term on the right of the equation, K(FG-1), as a description of the effects of the output variable on itself. It represents all the factors around the loop that lead from the output at time t, qo(t), to the new output, expressed as a change in output, qo(t+1)-qo(t). The multiplier of second term on the right, KF, describes only the organism mediated effects of input, expressed as (Hqd - qi*), on output. So it represents only the system (organism) factors in the loop that from input to output.

I propose that the efficiency of a control loop be defined as the ratio of the all factors contributing to system output, F(FG-1), relative to only the system's contribution (input) to the production of this output, KF. That is, I proposed defining the efficiency of a control system (CE for Control Efficiency) as

(3) CE = K(FG-1)/KF

which simplifies to

(4) CE = G - 1/F

This definition of efficiency makes a lot of sense to me. It says that the efficiency of a control system increases with increases in the size of the feedback function, G, relating output to input. It also says that increases in system amplification reduce efficiency, assuming F is negative (as it must be if loop gain, the product of G and F, is to be negative). For example, if G = 100 and F = -1 then CE is 101. If G = 100 and F = -100 then CE is 100.01.

Increasing G basically acts as a lever that magnifies the effect you have on the controlled variable. So if an organism controls in an environment that magnifies the effect of its output on controlled inputs (independent of the organism's own efforts) then the organism can control very efficiently (in terms of CE), in the sense that it has to contribute relatively little of it's own amplification, F, in order to control with high gain (the product FG is system gain and the larger FG is, in a negative direction, the better control is). So it's the amplification of the feedback function, G, that is the main determinant of the efficiency of control.

Two systems, with the same FG, will control equally well. But they could be quite different in efficiency. Assume that for one system F = -10 and G = 1000 and that for another system F = -100 and G = 100. The loop gain of both systems, FG, is the same: -10,000. However, the CE for the first system is 1000.1 while that for the second is 100.01. The first system is about 10 times more efficient than the second because it is operating in an environment that gives it 10 times the purchase, in terms of it's effect on the control variable.

Another way to look at it is like this: for the first system, a relatively small proportion of the energy for loop gain comes from F, which is the amplification provided by the system itself; for the second system, much more of the energy for loop gain comes from F. I consider the energy from the system itself as the system "input" to control loop's "output", which is the result of all sources of energy that contribute to loop operation. So the great G, the greater the control loop output per unit energy input.

I think the definition of control efficiency (CE) in equation 4 could provide an interesting basis for analysis of the evolution of control systems. It seems plausible that an evolutionary advantage would go to creatures that evolve to perceive and control those perceptual variables for which the particular environment they are in provides a high G feedback function. So if there are, say, three similar perceptual variables that an organism could perceive and control, the "successful" organisms -- the one's who end up successfully reproducing -- will be the one's that are designed to perceive and control the one out of those three perceptual variable to which the environment they are in -- their niche -- provides the strongest feedback connection, G.

Best

Rick

···

---

Richard S. Marken Consulting
marken@mindreadings.com
Home 310 474-0313
Cell 310 729-1400

[From Bjorn Simonsen (2006.07.06,21:20 EUST)]

From Rick Marken (2006.07.06.1040)

I propose that the efficiency of a control loop be
defined as the ratio

of the all factors contributing to system output,
F(FG-1), relative to

only the system’s contribution (input) to the
production of this

output, KF. That is, I proposed defining the
efficiency of a control

system (CE for Control Efficiency) as

(3) CE =
K(FG-1)/KF

which simplifies to

(4) CE = G - 1/F

Your words are also that “where G is a linear
multiplier that corresponds to the feedback function” and “F is a linear
multiplier that corresponds to the “system amplification function”.

I guess G has the units of “input per unit of output”
and F has the units of physical units per signal unit. (B:CP appendix)

The unit for CE is some complicated. And doesn’t the
slowing factor say anything about control efficiency?

I am in thinking modus and will remain there still
some time.

bjorn

[From Rick Marken (2006.07.06.1240)]

Bjorn Simonsen (2006.07.06,21:20 EUST)

Your words are also that �where G is a linear multiplier that corresponds to the feedback function� and �F is a linear multiplier that corresponds to the "system amplification function".

I guess G has the units of �input per unit of output� and F has the units of physical units per signal unit. (B:CP appendix)

The unit for CE is some complicated.

I think the units of CE output per unit input. This is based on the assumption that the units of G are input per unit output (i/o), as you say, and that the units of F are output per unit input (o/i). So the equation can be written:

CE = Gi/o - 1/Fo/i

Multiplying both sides by o/i I get

CEo/i = G - 1/F

But I don't trust my math or my dimensional analysis so I may be wrong. But it would be nice if CE were in units of o/i since this is how I think of efficiency: output per unit input.

And doesn�t the slowing factor say anything about control efficiency?

I thought it did but it cancels out given my way of defining efficiency. This makes sense if we assume that we are dealing with a nearly optimal controller. In such a controller, K is constrained to be 1/(1-FG) (equation 17 in Powwers' Psych Review paper). So in an optimal controller, when you know FG (which is what we assume we know when we measure CE) we know K (to be approximately 1/(1-FG).

I am in thinking modus and will remain there still some time.

Enjoy.

Best

Rick

Richard S. Marken Consulting
marken@mindreadings.com
Home 310 474-0313
Cell 310 729-1400

Both parasitism and symbiosis occur through reliance
on highly advantageous environments. Seeing these in
terms of efficiency for accomplishment (i.e.
perceptual control) makes sense to me.

Tracy

···

--- Rick Marken <marken@MINDREADINGS.COM> wrote:

...
which simplifies to

(4) CE = G - 1/F

This definition of efficiency makes a lot of sense
to me. ...

I think the definition of control efficiency (CE)
in equation 4 could provide an interesting basis
for analysis of the evolution of control systems.
It seems plausible that an evolutionary advantage
would go to creatures that evolve to perceive and
control those perceptual variables for which the
particular environment they are in provides a
high G feedback function. So if there are, say,
three similar perceptual variables that an
organism could perceive and control, the
"successful" organisms -- the one's who end up
successfully reproducing -- will be the one's
that are designed to perceive and control the
one out of those three perceptual variable to
which the environment they are in -- their
niche -- provides the strongest feedback
connection, G.

Best

Rick

__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around

[From Rick Marken (2006.07.06.2240)]

Tracy Harms wrote:

>Rick Marken wrote:

> ...
> which simplifies to
>
> (4) CE = G - 1/F
>
> This definition of efficiency makes a lot of sense
> to me. ...
>
> I think the definition of control efficiency (CE)
> in equation 4 could provide an interesting basis
> for analysis of the evolution of control systems...

Both parasitism and symbiosis occur through reliance
on highly advantageous environments. Seeing these in
terms of efficiency for accomplishment (i.e.
perceptual control) makes sense to me.

Nice example. Yes, parasitism and symbiosis seem to be adaptations that increase efficiency by using other organisms as part of an organisms' input control systems.

Best

Rick

···

---

Richard S. Marken Consulting
marken@mindreadings.com
Home 310 474-0313
Cell 310 729-1400

[From Bjorn Simonsen (2006.07.07,14:40 EUST)]

From Rick Marken (2006.07.06.1240)

I think the units of CE output per unit input. This
is based on the

assumption that the units of G are input per unit
output (i/o), as you

say, and that the units of F are output per unit
input (o/i). So the

equation can be written:

Your first sentence must be wrong.

CE = Gi/o

  • 1/Fo/i

Multiplying both sides by o/i I get

CEo/i = G - 1/F

But I don’t trust my math or my dimensional
analysis so I may be wrong.

But it would be nice if CE were in units of o/i
since this is how I

think of efficiency: output per unit input.

I think your dimensional analysis was correct in your
first mail and became wrong in your next mail after my mail. I now think I was
too quick with my first mail.

You mad a suggestion about CE in your [From Rick
Marken (2006.07.06.1040)].

I propose that the efficiency of a control
loop be defined as the ratio

of the all factors contributing to system
output, F(FG-1), relative to

only the system’s contribution (input) to the
production of this

output, KF. That is, I proposed defining the
efficiency of a control

system (CE for Control Efficiency) as

(3) CE
= K(FG-1)/KF

which simplifies to

(4) CE = G - 1/F

On your right side of the sign of equality in (4), you
present your definition. On your left side you have presented the name of your
variable “Control Efficiency”. At this moment your name on the left side must
have the same denomination as your definition on the right side.

Then you say:

So the equation can be written:

CE = Gi/o

  • 1/Fo/i

Multiplying both sides by o/i I get

CEo/i = G - 1/F

I think the first formula is OK. Here the name CE
already has the denomination “i/o”. This is what you presented in your first
mail. And now I agree when you said:

This definition of efficiency makes a lot of sense
to me. It says that

the efficiency of a control system increases with
increases in the size

of the feedback function, G, relating output to
input. It also says

that increases in system amplification reduce
efficiency, assuming F is

negative (as it must be if loop gain, the product
of G and F, is to be

negative).
For example, if G = 100 and F = -1 then CE is 101. If G =

100 and F = -100 then CE is 100.01.

Your last formula above doesn’t become “CEo/i = G -
1/F”. If you multiply with 1 o/I, CE becomes without denomination. And that is
also OK for me.

I agree with your first mail.

My reason for my comments was Bill’s first sentence
below Figure A-1 in his Appendix in B:CP. The ki (Your F) has units of signals
per physical unit and ko (your G) has units of (other) physical units per
signal unit. I had problems with “other” physical units.

I think your first mail was nice. I have learned
something new.

bjorn

···