Revised Cost of Schedule simulation

CSGnet Archive CSG1998
1

[from Jeff Vancouver 981103.1135 EST]

attached: Budget5.mdl

Attached is my latest rendition of my cost of schedule model. This version
incorporates two features not in the previous. First, to match the
blocking of the perception for trials 8 and 11, I have created a look-up in
"perceptual sensitivity" that mimics the blocking of the input in my
research paradigm. For some reason, the lookup is lagged by one trial
(that is, it is on trial 9 and 12 in the lookup table even though it is
trials 8 and 11 on which the effect occurs).

The second addition is for the goal given by the experimenter. To more
convincingly demonstrate the nature of control, I change the cost of
schedule goal for trials 3 through 6 in the research environment. To
parallel that in the model, I have added a control system that monitors the
goal from the environment. It never achieves control, but it does
determine the cost goal for each trial. Once again, I use a lookup to
accomplish this, and once again, it is lagged by 1 trial for some reason
(this is a detail I am not going to worry about).

That this last system never achieves control is simply an artifact of not
including the experimenter (who is pleased to some level) in the
environment. Perhaps I will work on that for the next rendition. In the
meantime, I have to figure out how to please the reviewers.

Sincerely,

JeffFrom ???@??? Tue Nov 03 20:34:53 1998
Return-Path: owner-csgnet@POSTOFFICE.CSO.UIUC.EDU
Received: from sj-mailhub-3.cisco.com (sj-mailhub-3.cisco.com [171.68.224.215]) by pilgrim.cisco.com (8.8.5-Cisco.1/8.6.5) with ESMTP id UAA07499 for <bnevin@pilgrim.cisco.com>; Tue, 3 Nov 1998 20:26:45 -0500 (EST)
Received: from proxy3.cisco.com (proxy3.cisco.com [192.31.7.90])
  by sj-mailhub-3.cisco.com (8.9.1a/8.9.1) with ESMTP id RAA19127
  for <bnevin@CISCO.COM>; Tue, 3 Nov 1998 17:42:17 -0800 (PST)
Received: (from smap@localhost)
  by proxy3.cisco.com (8.8.7/8.8.5) id RAA26316
  for <bnevin@CISCO.COM>; Tue, 3 Nov 1998 17:26:41 -0800 (PST)
Received: from postoffice.cso.uiuc.edu(128.174.5.11) by proxy3.cisco.com via smap (V2.0)
  id xma026284; Wed, 4 Nov 98 01:26:36 GMT
X-SMAP-Received-From: outside
Received: from postoffice.cso.uiuc.edu (postoffice.cso.uiuc.edu [128.174.5.11])
  by postoffice.cso.uiuc.edu (8.8.8/8.8.8) with SMTP id TAA56328;
  Tue, 3 Nov 1998 19:25:50 -0600
Received: from POSTOFFICE.CSO.UIUC.EDU by POSTOFFICE.CSO.UIUC.EDU
          (LISTSERV-TCP/IP release 1.8b) with spool id 10560416 for
          CSGNET@POSTOFFICE.CSO.UIUC.EDU; Tue, 3 Nov 1998 19:25:48 -0600
Received: from mtiwmhc01.worldnet.att.net (mtiwmhc01.worldnet.att.net
          [204.127.131.36]) by postoffice.cso.uiuc.edu (8.8.8/8.8.8) with ESMTP
          id TAA92662 for <CSGNET@POSTOFFICE.CSO.UIUC.EDU>; Tue, 3 Nov 1998
          19:25:47 -0600
Received: from fnickols ([12.78.156.220]) by mtiwmhc01.worldnet.att.net
          (InterMail v03.02.03 118 118 102) with SMTP id
          <19981104012516.KUIX9083@fnickols> for
          <CSGNET@POSTOFFICE.CSO.UIUC.EDU>; Wed, 4 Nov 1998 01:25:16 +0000
X-Sender: nickols@postoffice.worldnet.att.net (Unverified)
X-Mailer: QUALCOMM Windows Eudora Light Version 3.0.5 (32)
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Message-ID: <3.0.5.32.19981103202728.0082f100@postoffice.worldnet.att.net>
              <CSGNET@POSTOFFICE.CSO.UIUC.EDU>
Sender: "Control Systems Group Network (CSGnet)"
              <CSGNET@POSTOFFICE.CSO.UIUC.EDU>

budget5.mdl (60 Bytes)

···

Date: Tue, 3 Nov 1998 20:27:28 -0500
Reply-To: "Control Systems Group Network (CSGnet)"
From: Fred Nickols <nickols@WORLDNET.ATT.NET>
Subject: Letter to the Editor (HBR)
To: Multiple recipients of list CSGNET
              <CSGNET@POSTOFFICE.CSO.UIUC.EDU>
X-UIDL: ffc3e0e88435622b279caa700aec4c46

I just received the Nov-Dec issue of Harvard Business Review.
The letter I shared with the list a while back is in fact there.

Boy, do I feel good.

Regards,

Fred Nickols
Distance Consulting
http://home.att.net/~nickols/distance.htm
nickols@worldnet.att.net
(609) 490-0095

2

[From Bill Powers (981111.0727 MST)]

Jeff Vancouver 981103.1135 EST--

attached: Budget5.mdl

Attached is my latest rendition of my cost of schedule model. This version
incorporates two features not in the previous. First, to match the
blocking of the perception for trials 8 and 11, I have created a look-up in
"perceptual sensitivity" that mimics the blocking of the input in my
research paradigm. For some reason, the lookup is lagged by one trial
(that is, it is on trial 9 and 12 in the lookup table even though it is
trials 8 and 11 on which the effect occurs).

The second addition is for the goal given by the experimenter. To more
convincingly demonstrate the nature of control, I change the cost of
schedule goal for trials 3 through 6 in the research environment. To
parallel that in the model, I have added a control system that monitors the
goal from the environment. It never achieves control, but it does
determine the cost goal for each trial. Once again, I use a lookup to
accomplish this, and once again, it is lagged by 1 trial for some reason
(this is a detail I am not going to worry about).

That this last system never achieves control is simply an artifact of not
including the experimenter (who is pleased to some level) in the
environment. Perhaps I will work on that for the next rendition. In the
meantime, I have to figure out how to please the reviewers.

Are you aware that the behavior of this model depends critically on the
Time Step you set up in "Models ... Time Bounds"? Try setting it to 0.1!

A properly set up model should give exactly the same behavior as long as
the time step is short enough: the run will take longer if the time step is
shorter, but the graphs should look the same. The present model is not
independent of the size of the time step.

The basic problem with this model is that the variables and relationships
are just thrown together with no justification or logic. Looked at
numerically, what you have on the first trial is a goal cost of $14000
times a goal sensitivity (whatever that means) of 1, minus the value of
"please experimenter" which is 1, for a grand total "goal error" of 13999,
which becomes the "cost goal", also 13999. That cost goal is logically
compared with perception of cost, which is the cost of schedule times a
perceptual sensitivity near 1.

I'm beginning to appreciate Bob Eberlein's preference for using the names
of the external variables represented by signals instead of "NU" (nervous
system units). When you use NU for everything inside the model of a brain,
you can do such things as subtracting one unit of desire to please an
experimenter from a signal representing $14000, to get 13999 units of
something or other. While that could be done with signals, it makes no
sense in terms of what the signals mean in the outside world.

You did so well with the thermostat model: why did you get so sloppy with
this one? I urge you not even to THINK of publishing this model.

Best,

Bill P.

3

[from Jeff Vancouver 981111.1130 EST]

First let me say that I am pleased you are interested enough in this
project to give it your time. Second, let me see if I understand your
problems with the model.

[From Bill Powers (981111.0727 MST)]

Are you aware that the behavior of this model depends critically on the
Time Step you set up in "Models ... Time Bounds"? Try setting it to 0.1!

I was not aware of this. Now I am. I do not understand it.

A properly set up model should give exactly the same behavior as long as
the time step is short enough: the run will take longer if the time step is
shorter, but the graphs should look the same. The present model is not
independent of the size of the time step.

I am thinking that what you mean by this is that a properly set up model
includes some understanding of the physical lags involved. Hence, if it
takes a second for some change to occur, it takes a second for that change
to occur no matter what the time step of the simulator (as long as the time
step is <= to 1 second). I have made no attempt to parameterize the lag of
the physical system. I suspect that as a result of that, I am getting
strange results when the time step is changed (in either direction). Can
you help me with this?

The basic problem with this model is that the variables and relationships
are just thrown together with no justification or logic. Looked at
numerically, what you have on the first trial is a goal cost of $14000
times a goal sensitivity (whatever that means) of 1, minus the value of
"please experimenter" which is 1, for a grand total "goal error" of 13999,
which becomes the "cost goal", also 13999. That cost goal is logically
compared with perception of cost, which is the cost of schedule times a
perceptual sensitivity near 1.

First of all, the "please experimenter" constant is 0, but that does not
change your basic point. Second, the sensitivity constants are there to
change the units from external units (e.g., $) to internal units (NU). I
copied that from the Thermo models.

I'm beginning to appreciate Bob Eberlein's preference for using the names
of the external variables represented by signals instead of "NU" (nervous
system units).

And I, because of this model, appreciate your preference.

When you use NU for everything inside the model of a brain,
you can do such things as subtracting one unit of desire to please an
experimenter from a signal representing $14000, to get 13999 units of
something or other.

Exactly.

While that could be done with signals, it makes no
sense in terms of what the signals mean in the outside world.

Yes, Bill, this is the one of the beauties of your theory. The meaning is
completely determined by the configuration of the system. In the past I
have asked where the system-level reference signal comes from (the highest
in the hierarchy). Rick said that did not bother him, it could be just 0,
or because of that, not there. Indeed, if we assume a simple human, the
"please experimenter" could represent that level (at least this is the
system value I want all my undergraduate participants to have as their
higher-level goal). Seriously, I could have done just what Rick suggested,
that is, not include it and just have the perception of goal go directly to
cost goal. I did not because I think that leads to misunderstandings
regarding control. However, my point is that the nature of that highest
level reference signal (at least in my model) need not be in the units of
the external variables that are eventually controlled as a result of
discrepancies from it.

You cannot argue that humans are internally wired to think in monetary
units. The beauty of your model (and the use of NU) is that you do not
have to. The system can control using its own units.

Now it is true that we learn things about external units that constrain and
inform the types of perceptions and output functions we create to control
CVs with those units. But as you told Bob, dealing with that level of
meaning and its impact on the systems is a very difficult problem we are
not ready to tackle yet. I am making some assumptions in creating my model.

This is also partially true of the timing issue. Lag is important and many
of you have learned to deal with it in your modeling. I have not yet and
so assumed it away. I think you can correct me on that. On the issue of
the meaning of the units and how that meaning is represented internally,
that is more than I suspect even you can handle. But using the constaints
of the physical world as a proxy for constaining the internal processing
seems to me a mistake that will eventually get you into trouble. Let me
reiterate, I think your original position with Bob was the correct one.
Perhaps you have a better way to model the change in the goal so that
somehow the adoption of this externally presented goal becomes an
internally represented goal. Some way that does not ever leave the units
as dollars. But I am not sure why you feel that is necessary.

You did so well with the thermostat model: why did you get so sloppy with
this one? I urge you not even to THINK of publishing this model.

Now I am conflicted. I am trying to control the perception that others
perceive me with respect, so I do not want to publish something that makes
me look like a fool. Meanwhile, I am trying to show scientists how we can
study control systems (and get grant money, tenure, etc.). I may not
publish this model, but I certainly want to publish a version of it.

I am not sure if there are other simplifications that I am making that are
bothering you. Not including the experimenter as an external variable, or
having the perception of goal input function contain the mechanics of the
computer program that changes the goal for trials instead of having the
operation occur in the environment before becoming input to the input
function are two such simplifications. I just noticed that I called a
sensitivity constant "goal gain" by mistake. These are details that need
to be fixed and may lead you to say "that the variables and relationships
are just thrown together with no justification or logic."

Indeed, I am seeking simplicity lest I lose the audience. It would be much
simplier to remove the sensitivities and gains and hence all the unit
changing that they allow. But I think that would be a mistake.

Sincerely,

Jeff

4

[From Bill Powers (981112.0226 MDT)]

Jeff Vancouver 981111.1130 EST--

I'm off in a few hours to Charlotte NC until Sunday, so can't do much with
your model right now. I think the problems can be at least reduced, by
narrowing the range of phenomena that you try to cover with the model
("pleasing the experimenter" is a bit too ambitious). Instead of trying to
model "trials" with a kludge involving integrating 1's (this is what makes
your model's behavior dependent on the size of the time step), you need to
detect events such as the cost being reduced to or below the criterion
level and use those to initiate substituting a new cost. This way you don't
get tied to the time scale of the simulation.

Later

Bill P.

5

[from Jeff Vancouver 981116.1450 EST]

[From Bill Powers (981112.0226 MDT)]

I'm off in a few hours to Charlotte NC until Sunday, so can't do much with
your model right now. I think the problems can be at least reduced, by
narrowing the range of phenomena that you try to cover with the model
("pleasing the experimenter" is a bit too ambitious). Instead of trying to
model "trials" with a kludge involving integrating 1's (this is what makes
your model's behavior dependent on the size of the time step), you need to
detect events such as the cost being reduced to or below the criterion
level and use those to initiate substituting a new cost. This way you don't
get tied to the time scale of the simulation.

I trust NC was fun.

I am trying to figure out your suggestion. I do have a control system that
detects the event of the cost of schedule dropping below the goal. The
issue is having trial as a static variable (it increments when the event is
detected). I would love a way. Must be one. I will keep looking. Of
course, this would be easy in a conventional programming language.

Sincerely,

Jeff