Professional Documents
Culture Documents
was not the correct one (i.e., the one spends more or less randomly. This
assigned to the stimulus presented on means that the function given in the
that trial); at other times more elabo- above formula necessarily levels off at
rate methods of "correction" are used some positive asymptote for large
(26). In any case the so-called as- interstimulus distances. Again, in
signment is simply a rule which the order to compare data from different
experimenter follows in delivering the experiments, this asymptote must be
reinforcements, and any arbitrary brought down to zero for each experi-
one-to-one assignment can be estab- ment. This is accomplished by esti-
lished in this way. mating a parameter Cs from each set
Consider then a paired-associate of data (25), and by defining the gen-
experiment in which the responses are eralization between Si and 5* to be
highly distinctive and so lead to neg-
ligible amounts of generalization. If
there are N pairs consisting of one P;SP,
After removal of the stimulus, the dis- which does not depend upon the choice
tribution of trace elements progres- of an arbitrary interval At. Then, if
sively spreads out in this space, as the <-axis is translated so that condi-
illustrated in D, E, and F. tioning takes place at t = 0, m can be
In addition to this spread or diffu- increased in such a way that, as At
sion of the trace, it is assumed that the 0, m- At -t.
trace elements are also subject to Although the probability that a
spontaneous deconditioning (again, given trace element is deconditioned at
presumably owing to random proc- precisely time t is zero, the probability
esses at the molecular level). Thus, per unit time (the "probability den-
as the elapsed time increases, the sity") or rate of deconditioning for an
number of elements still conditioned individual trace element is, at the
decays to zero. In Fig. 3, the shaded particular instant t,
areas represent the fraction of the
original trace elements which are still lim [1 - U(Af)Ji^U(At)/At
A(->0
conditioned at each time.
But by Equation 5 as At -* 0, U(At) >
The Deconditioning of the Trace U-At+Ut/m. Therefore, the rate
The probability that a given condi- of deconditioning at time / is, for
tioned element will suffer decondi- single trace elements,
tioning during a small interval At will lim [1 - Ut/m~]mU
be denoted by U(At). The simplest m*oo
rule that can be assumed to govern = 7 exp (- 70 [6]
this probability is as follows:
Assumption I. U(At) is a constant, where, for convenience in what fol-
independent of the time the element lows, exp ( Ut) is used in place of
has remained conditioned and inde- e~ut.
pendent of the distance (in psycho- With regard to the deconditioned
logical space) to which the element has trace elements, the following rule is
drifted in that time. the simplest that can be assumed:
The probability that an element, Assumption II. In the absence of
still conditioned at time /, remains further reinforcements, a decondi-
conditioned until t + At is, of course, tioned element remains deconditioned.
1 - [/(Aif). Therefore, the probabil- From this assumption and Equation
ity that a given trace element remains 6 it follows that the fraction of the
conditioned during the first m intervals originally conditioned elements re-
of length At, but then becomes decon- maining conditioned at time t is
ditioned during the immediately suc-
ceeding A/-interval, is, by Assumption U(f) = 1 - f ' 7 exp (- Ur)dr
I,
[1 - U(AO~]mU(At) [4] = exp (- 70 [7]
Now U(At) is necessarily propor- The Diffusion of the Trace
tional to the interval chosen for At.
It is convenient, therefore, to define Equation 7 completes the quantita-
a deconditioning parameter, tive formulation of the deconditioning
of the trace. In order to provide a
7(AQ similar formulation for the diffusion
U = lim [5]
A(->0 At of the trace, it is necessary to consider
STIMULUS AND RESPONSE GENERALIZATION 249
the motions of the individual trace gible. This can be stated with greater
elements in psychological space. The precision as follows :
exposition is simplified by continuing Assumption IV. The function f&t
to suppose that the stimuli are evenly and, thus, the probability density of
spaced along a restricted range of a displacements of a trace element from
single physical dimension. It is then x = 0 is distributed over the #-axis
possible to introduce one coordinate with finite variance.
Xi for each stimulus representation Si* From Assumptions III and IV it
giving its position along a one-dimen- follows that the variance of the dis-
sional psychological space. The psy- tribution of displacements during an
chological distance between any two interval At is the finite constant
stimuli, Si and Sk, is then
Da, = \Xi xk\
V(Af) = [11]
[8]
Now the expression Vik(Ax, At) Since V(Af) must depend upon the
will denote the probability that, if at length of the interval At, it is useful to
time t a trace element is situated at Xi, define a diffusion parameter,
by time t + At it will have moved into
the one-dimensional region bounded V = lim [12]
by Xk and Xk + Ax. As before, the Af-0 At
arbitrary interval Ax can be eliminated which does not require the stipulation
by defining a new quantity of an interval At. Just as U governs
the rate of deconditioning of the trace,
[9] then, V controls the rate of spread or
diffusion of the trace in psychological
is the probability density space.
of displacements from Xi to Xk during The next question to be answered is
the brief interval A^. The simplest this : Given the diffusion parameter V,
rule that can reasonably be assumed what form will the distribution of
to govern this quantity is as follows: trace elements take after an appreci-
Assumption III. Vik(At) is an able delay t? It is possible to show
invariant function of the psychologi- that the assumptions which have been
cal distance between 5,- and Sk. It made are sufficient conditions for the
does not depend upon the time ^ or desired distribution to tend toward a
upon the absolute position of the pair, limiting form which is independent of
Si and Sk, in psychological space. the form of /A ( (6, 15). Specifically,
If the x-axis is translated so that # if a trace comprising a large number
= 0, it follows from Assumption III () of elements is conditioned from Si
that, for a given interval At, there to Xi at t = 0, the density of these
exists a fixed function /A t, such that elements at x/, for some later time t
conforms with the Gaussian function
Vik(At) - [10]
n-Vik(t) = n-(2ir7/)-*
However, it will not be necessary to X e x P [ - (A [13]
make any particular assumption about
the form of the function /A*. It is The beauty of this result is its com-
only necessary to insure that, during plete independence from the under-
a short period of time, the probability lying mechanism symbolized by f&t-
of a very large displacement is negli- Thus, even though one imagines, for
250 ROGER N. SHEPARD
example, that thermal agitation of the initial phase of rapid learning, rein-
molecular substrate is responsible for forcements occur at relatively fre-
deconditioning and diffusion, the bio- quent and regular intervals. If the
physical details of this process need early trials are disregarded, then, a
not be specified. For, according to roughly steady-state process can be
the present formulation, these details considered. The summation of the
are irrelevant to the question of the Gaussian curves of various ages can
gross behavior of the trace system. then be approximated by an integra-
tion of these curves over t.
The Trace Process in Paired-Associate Now, with respect to Si, the density
Learning of trace elements at Xk resulting from
a reinforcement t time units ago is
Now in the course of learning paired
given by n- Vik(t). However, only
associates, each stimulus Si will have the fraction U(t) of these is still con-
been presented on many occasions.
ditioned. Therefore the density of
Furthermore, on a number of these
elements at Xk which remain condi-
occasions, reinforcement of the re-
tioned after a delay t from reinforce-
sponse assigned to St will have condi-
ment is n U(t) Vii, (f) Clearly, then,
tioned a bundle of n trace elements to
the total density of conditioned trace
Si*. Thus at some given time /o, the elements resulting from all previous
density of conditioned elements re-
reinforcements of the response assigned
sulting from the immediately preced-
to Si is distributed in approximate ac-
ing reinforcement will be distributed
cord with
in psychological space as shown for
t-i in Fig. 4. Likewise, the densities
remaining from earlier reinforcements (X n-U(t)-Vik(t)dt
Jo
will be distributed as shown for L.2,
2_3, and so on. c
= I w-exp ( Ut) u.TrV()~*
The total distribution of condi- Jo
tioned elements emanating from Si at exp [- (Dik)i/2Vf\dt [14]
to can be found by summing the
Gaussian distributions resulting from Fortunately, the integration can be
all previous reinforcements of the re- effected (4, p. 144) and, indeed, yields
sponse assigned to S,. It is possible to
derive an analytic approximation to n(2UV)~
this composite distribution (J"t in [15]
Fig. 4) by supposing that, after an
For intermediate stages of paired-
s* associate learning, then, this function
can be taken as a measure of the
strength of connection of the stimulus
Si to the point xk on the perceptual
continuum. It is not a probability
density, however, since (if n> 1 or
U < 1) the integration of this func-
tion over the x-axis yields a value
FIG. 4. A series of Gaussian distributions greater than unity (7, p. 133).
with increasing variances, and the composite
distribution arising from an integration of Multiplication of Equation 15 by
these over time. U/n converts this function to a prob-
STIMULUS AND RESPONSE GENERALIZATION 251