Order reduction; More VI data

[From Bill Powers (950717.1410 MDT)]

Bruce Nevin (950717.13:32 EST)--

     In particular, I believe you don't mean to say the following:

     When you use a word, such as "jump", to name a perception, you
     create a link of some kind between that word-perception and the
     perception that you name by it. Once you have created this link,
     either the word-perception or the other perception can evoke a
     category perception.

Right, this is not what I mean (I think -- my thinking about this
subject is not what you would call advanced).

As described above, the process is simply the naming of a specific
perception: associating one word-perception with one nonverbal
perception. If this were all that is involved in naming, we would soon
run out of words.

A category perception, as I imagine it, is a signal that is generated
upon the occurrance of any of a set of specific nonverbal perceptions at
a lower level. Words are not involved. Controlling a category perception
means producing any of the experiences of lower level that are treated
as members of the same category. For example, perceptual items that have
proven to be tasty and nonpoisonous in the past are useful to lump into
a category because eating any of them will prove at least nourishing --
will alleviate hunger without harm and perhaps with pleasure. An animal
without language would be able to use such a control system as a way of
discriminating edible items when it is hungry, even though it had no
name for such a category of items, and even though the particular edible
item that will be found is not predictable.

This mode of perception might have something to do with our ability to
pick out an odd item in a collection of similar ones:


All of the alphabetic configurations belong to one perceptual category;
the 3 belongs to a different one. So there are only two category signals
involved, no matter how many letters there are. This makes the 3 easy to
find. The 3 "jumps out" -- it doesn't seem to me that there is an
intermediate stage of saying


In other words, I don't think that naming of the categories is involved.

Naming a category is including a word-configuration in a category
otherwise composed of non-word items.

                                      > >
            word configuration ------>|CATEGORY|
                                      >INPUT | -->Category signal
            nonword items ----------->|FUNCTION|
                                      > >

So either the word or one of the nonword items would create the sense of
the same category. "DOGS" or an image of any dog would evoke the
impression of the category dogness, itself a nonverbal signal. Note that
there is no direct link between the word configuration and the nonword
items that belong to the same category. All that they have in common is
that they evoke the same specific impression of category-ness.

Anyway, that's how my concept of naming categories works. I don't know
how it will stand up to criticism.

P.S. When you CC to me as well as sending to CSG-L, I get the same post
both ways. The post to CSG-L is enough. For private posts you can use my
address alone.


Bruce Abbott (950717.1405 EST) --

Very complicated data! When we look at dynamic considerations, we will
need a more complex model. In my old operant conditioning model, where I
had an explicit perceptual function integrating the cumulative
individual reinforcements, an effect like the ones in these data would
appear. After a long period of deprivation, the perceptual signal would
have decayed essentially to zero. So when behavior became effective,
there would be a high rate of responding at first, while the large
deficit was made up, followed by a decline to an equilibrium state that
would go on indefinitely. This might account for the change in shapes of
the curves with temporal position in a session.

As you suggest, this might be a two-level problem. On the other hand, I
feel that the second level would have to operate too fast, unless
there's something markedly different about controlling for water and
controlling for food. Worth a try, anyway.

This is a worthy candidate for advanced modeling, but I'd like to see us
tackle something simpler, first. In at least one earlier set of data,
you reported that the authors waited until the variations in behavior
were less than 5% before starting to take real data, just to make sure
that asymptote had been reached. The model we're working with now would
be applicable to that kind of data. Looking at the dynamics of control
at 5-minute intervals requires something more than a steady-state model,
and I'd like to put that off.

Interval 2nd 3rd 4th 5th
Rft/hr drops ( ml ) drops ( ml ) drops ( ml ) drops ( ml )
240 20.0 ( 5.0) 40.0 (10.0) 60.0 (15.0) 80.0 (20.0)

You comment that "by the start of the fifth interval on the 240
rft/hr (VI 15-s) schedule the rat has consumed up to 20 ml of water (20
minutes * 4 rft/min * 0.25 ml/rft)". It looks to me as if the rat has
consumed 5 + 10 + 15 = 30 ml of water by the start of the 5th session,
plus whatever it consumed in the first interval (why wasn't that
shown?). Unless the numbers you're showing above are cumulative???

Also, you say "In contrast, by the same interval on the 15 rft/hr (VI
240-s) schedule it has only consumed 0.1 ml at best." Adding up the
numbers for sessions 2,3 and 4, I see 1.8 ml total consumption. Better
explain more fully. Or check your arithmetic. Seems to me a spreadsheet
might be a good way to fiddle with these numbers.

We really would need the obtained reinforcements.
     Another consideration to keep in mind is that responding does not
     go to zero in the absence of reinforcement; rather it goes to some
     "baseline" level which is the product of exploratory activity etc.
     The low rates of responding given for the VI 240-s schedule appear
     consistent with baseline levels (although no baseline rates are
     presented in the report).

That's a good point: even if the behavior rate on the specific task had
actually dropped to zero, the animal would periodically revisit the key
or bar just in the course of trial-and-error pecking or pressing, and we
would see a non-zero behavior rate. Another reason for keeping track of
where the animals are.
     Studies reporting average rates across sessions would be
     contaminated by this effect. As yet I don't know whether this
     effect is as pronounced when the typical 45 mg food pellets are
     used as reinforcers, but I believe I can find data which can be
     used to check.

The degree of contamination would depend on how long the total session
was. If the changes had essentially reached asymptote after the 5th 5-
minute interval, averaging over 12 intervals wouldn't create too great
an error.

A pertinent paper (picked up from JEAB abstracts on WWW from a search on
"reinforcement theory":

Hunter, I., & Davison, M. (1985). Determination of a behavioral transfer
function: White-noise analysis of session-to-session response-ratio
dynamics on concurrent VI-VI schedules. JEAB _43_, 43-59.

They showed a transfer function with a gain of 0.53 and a time constant
of 0.67 sessions -- unfortunately the abstract didn't state the duration
of a session. The time constant related to the time needed to reach
asymptote after a switch in schedules, with about 5 sessions giving 96%
approach to the asymptote. They found the gain of 0.53 surprising, as it
didn't jibe with the sensitivity to reinforcement ratio under steady-
state conditions; they probably made a mistake, or were looking at the
second level of control.

Anyway, it would be interesting to see if their actual time-constants
agree with those in the experiment you cite here.
Best to all,

Bill P.