Gibbon 1977
Gibbon 1977
Gibbon 1977
The rapid growth of research that fol- shaped key pecks occur (Terrace, Gibbon,
lowed Brown and Jenkins's (1968) demon- Farrell, & Baldock, 1975). When / is long,
stration of autoshaping has sharpened the fewer pairings of the key light and food
need for delineating the nature of this phe- are needed before a peck occurs than when
nomenon and its relation to traditional con- / is short. This result is comparable to the
ditioning paradigms. A recent study per- frequent (though not universal) finding in
formed in the authors' laboratory has shown classical conditioning paradigms of more
that the intertrial interval (/) between key rapid and stronger conditioning with spaced
illuminations exerts a powerful effect upon than with massed training (e.g., Gorme-
the number of trials required before auto- zano & Moore, 1969).
At present, there are only fragmentary
data concerning the joint contribution of in-
These experiments were supported by National tertrial interval and trial interval (T) dura-
Science Foundation (NSF) Grants GB 34095 and
BNS 76-01229 and Small Grant MH 2S070 to tions. Perkins et al. (1975) suggest that
J. Gibbon and by NSF Grant GB 30781 and Na- long trials result in lower rates of pecking
tional Institutes of Health Grant HD 00930 to than do short trials. That study also con-
H. S. Terrace. firmed the inverse relationship between I
The data reported in Experiment 1 were origi- duration and trials to acquisition reported
nally presented in the 1975 Columbia University
doctoral dissertation of M. D. Baldock, "The by Terrace et al. (1975). However, Perkins
Effects of Trial and Intertrial Interval Durations et al.'s study focused mainly on the main-
on the Acquistion and Maintenance of the Auto- tenance of "autopecks" and not on their
shaped Key Peck." These data were also partially acquisition.
reported at the Eastern Psychological Association
meeting in Philadelphia, April 1974. Brown and Jenkins's (1968) study sug-
M. D. Baldock is now at the Australian Na- gests joint effects of trial and intertrial du-
tional University, Education Research Unit, The ration. In one of their groups, the key light
Research School of Social Sciences, Box 4 PO, was on continuously throughout training,
Canberra, A.C.T. 2600, Australia. and the hopper appeared at the same inter-
Requests for reprints should be sent to J. Gibbon,
New York State Psychiatric Institute, 722 West vals as were used for the autoshaping group
168 Street, New York, New York 10032. in which the light preceded food and was
264
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 265
extinguished in the intertrial interval. No chambers with the centered key. Accordingly, no
pecking occurred in the continuous group. further distinction will be made between the two
types of chambers.
Such a group might be thought of as analo- A force of approximately 18 g was required to
gous to one in which the intertrial interval displace each of the response keys to interrupt the
is very short and the trial occupies virtually normally closed contact at the back of the key.
all of the period between reinforcements (7 Each contact closure provided a brief "feedback"
— 0). The only conditioning that is possible click. Each chamber was equipped with a 1S-W
houselight mounted in the roof. The houselight
here would be temporal conditioning to the was illuminated continuously except during hopper
spacing of reinforcements. activation when a hopper light was the only illu-
At the other extreme, one may consider mination. Under all conditions, a trial was defined
a situation in which the intertrial duration by the interval during which the response key was
transilluminated by a green light. This stimulus
occupies most of the period between food was generated by an IEE projection device (Model
presentations and the key light comes on #10M54-Q-44B) which directed the output of a
very briefly before or nearly simultaneously 28-V bulb (GE 1828) through a green Wrattan
with reinforcement (T — 0 ) . Temporal con- filter onto the rear of the response key. The lumi-
nance of the homogeneous field of green light was
ditioning would again seem to be the only -1.2 log ftL.
basis for associative effects here (cf. Stad- Procedure: Hopper training. Prior to auto-
don & Simmelhag, 1971). shaping, each subject was required to satisfy an
The present experiments examined the eating criterion. At the start of hopper training,
contribution of I and T values between these a naive subject was placed in an experimental
chamber in which the response key had been cov-
extremes. Experiment 1 studied fixed trial ered. The only source of illumination in the cham-
and variable intertrial durations. Experi- ber was the hopper light. Prior to placing the bird
ment 2 studied fixed trial and fixed inter- in the chamber, a few grains of food were scat-
trial durations. In both experiments, an at- tered on the floor immediately in front of the hop-
per aperture. Typically, subjects ate this grain al-
tempt was made to encompass a substantial most immediately and subsequently ate from the
range of values of both variables and to in- food hopper. Whenever the subject inserted its
clude combinations for which the ratio of head into the hopper, it interrupted the beam of
these two parameters was constant. light directed at a photocell; 3% sec later, the
hopper was lowered and the houselight illuminated
the chamber. Following an average interval of 20
Experiment 1 sec, the hopper was raised again.
Upon completion of 10 eating trials, the first
Method session was terminated after the first sequence of
5 trials with eat latencies less than or equal to
Subjects. The subjects were 136 male and fe- 1.5 sec. If no such sequence occurred, the session
male white Carneaux pigeons. At the start of the terminated after 50 eating trials. Every subject
experiment, each pigeon was experimentally naive. was required to satisfy the latency criterion twice
Their ages ranged from 6 to 9 months. Through- during no more than four hopper training sessions.
out the experiment, subjects were maintained at Subjects which did not do so were excluded from
80% ± 3% free-feeding weight. the experiment. This screening resulted in an at-
Apparatus. Four experimental chambers were trition rate of about 20%.
used. Each was made from modified Coleman ice Procedure: Autoshaping. Figure 1 shows the
chests with internal dimensions of 30 cm wide X experimental design on double log coordinates.
31 cm high X 31 cm long. Three of the chambers The / values are shown on the #-axis and T
had the response key located behind a circular values are shown on the y-axis. Combinations with
aperture in the middle of the front panel, directly constant ratios are those with a common positive
above the hopper opening. The fourth chamber was diagonal. In each case, T was fixed and / repre-
originally a two-key chamber. One of these keys sented the mean of an approximately exponentially
was covered (and painted flat black) throughout distributed sample of intertrial intervals. The dis-
the experiment. In the fourth chamber, the opera- tribution for I = 96 sec is given in Appendix A.
tive key was located approximately 7 cm to the The other / distributions may be obtained from
left of the hopper opening. In all chambers the this by multiplying all the intervals by the appro-
center of the response key was 20.32 cm above the priate constant (e.g., all intervals are halved to
floor. A pilot study demonstrated that there were obtain the 48-sec distribution).
no differences between the acquisition and per- After satisfying the magazine criterion, birds
formance of subjects conditioned in the chamber were randomly formed into groups and studied in
with the offset key and subjects conditioned in the sessions lasting for 25 trials. \Sessions began with
266 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
an intertrial interval of a duration that was the from these groups were subjected to analysis
mean value of the variable sample programmed of variance. The measure used was the log
for that bird for that condition. Thus, for exam-
ple, birds in the groups at / = 96 sec were ex- of the number of trials subjects experienced
posed to an intertrial interval of 96 sec at the prior to satisfying an acquisition criterion
beginning of each session. Subjects were studied of a sequence of three out of four trials in
either for 20 sessions without a response or for a row with at least one response. This mea-
20 sessions beyond the session in which they first
began to peck the key. sure reflects the beginning of sustained re-
sponding and is less variable than trials to
the first peck. The log transform was used to
Results
reduce an otherwise substantial correlation
Acquisition. Acquisition results con- between mean and standard deviation. Tests
firmed and extended our earlier findings of homogeneity of variance on the logged
(Terrace et al., 1975). The earlier study data (Pearson & Hartley, 1970, Table 31,
found trials to acquisition to be a negative interpolated for k = 20) approached but did
power function of / value. In the present not reach significance. Thus, the log trans-
study, this relationship was found to hold form reduced inhomogeneity.
within each T value. However, the trial du- The analysis of variance confirmed a very
ration modulated the overall level of acquisi- large effect of intertrial interval, F(4, 85) =
tion. When T was long, acquisition was re- 10.23, p < .001, and also a reliable effect
tarded, and when T was shortened (down of trial duration, F(3, 85) = 2.85, p < .05.
to T = 2 sec), acquisition was facilitated. The interaction between these two variables
One group of four subjects, studied at the was not significant, indicating that the log
combination I = 96 sec, T = 1 sec, did not transform resulted in parallel effects of the
acquire key pecking. Acquisition data on in- intertrial interval at different trial values.
dividual subjects is contained in Appendix Group medians of individual subject's
B. Four trial durations (T = 4, 8, 16, and number of trials to acquisition are shown in
32 sec) were studied at a broad range of Figure 2 as a function of intertrial interval
intertrial interval values (/ = 24, 48, 96, with trial duration as a parameter. Both
384, and 768 sec), and acquisition data axes are logged. The four functions shown
in the figure are least squares fits and there-
fore describe a power relationship between
1.5 2 3 acquisition and / values. The correlations on
individual functions all exceeded .98, and
64
the associated F for regression (cf. Draper,
-5 32 1966, pp. 38-41, and Ergun, 1956) is ac-
cordingly very large, F(l, 9) = 340.0, p <
.001. Individual slopes were found not to
differ reliably from each other and are well
represented by the pooled estimate of $ =
— .777. Evidently, changing trial duration
alters only the level of these functions. At
a given trial duration, increasing the inter-
e 24 48 96 384 768 trial interval facilitates acquisition at about
INTERTRIAL INTERVAL (I in sec) the same rate as increasing the intertrial in-
Figure 1. Experimental design: Entries in the
terval by the same factor at a different trial
graph are the number of subjects studied at the co- duration.
ordinate combination of intertrial interval (I, ab- Changing trial duration changes the in-
scissa) and trial interval (T, ordinate) values. tercepts of these functions, so that the pro-
(Both axes are logged so that constant ratios of portionality constant in the underlying
I IT [indicated along the top of the figure] result
in diagonal lines. Groups with a common ratio are power function relating the two variables
those connected by these diagonals.) depends on T. This dependence is shown in
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 267
T — 32 sec do not have the same character and there was a substantial effect for trial
as those for the Group / = 24 sec, T = 16 duration, F(3, 47) = 10.58, p < .001. The
sec. Similarly, compare 1 — 96 sec, T — 4 interaction approached but did not reach
and 8 sec with / = 384 sec, T = 16 and 32 significance, F(9, 47) = 1.84, .05 <p< .10.
sec. However, the missing data technique
The overall mean response rate in the (Winer, 1971, p. 48) was required in this
trial for each subject over the final 4 days of analysis, and this tends to minimize inter-
training was logged and the data subjected action effects.
to an analysis of variance. The analysis The trial duration effect is shown in Fig-
confirmed the general findings noted above. ure 6 on double log coordinates. The best fit
There was no effect for intertrial interval, power function, indicated by the line, ac-
T=4 T=8 T-16 T=32 AVERAGE
1=24
1=48
1=96
1=384
1=768
3 5 1 3 5 1 3 5 I 3 5
FIFTHS OF THE TRIAL
Figure S. Response rate on a log scale in successive fifths of the trial signal for selected groups.
Intertrial (/) values increase from top to bottom of the figure, and trial (T) values increase
from left to right. (Xs are for the 1st day of responding [with at least 10 trials responded to],
open circles are for the first 4 days of responding, and filled circles for the last 4 days of
responding.)
270 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
4 10 20 40 4 10 20 40
TRIAL DURATION
Figure 8. Median latency (in seconds) of the first response in the trial as a function of trial
(T) duration (in seconds) for the 1st and the 15th sessions (left and right panels, respec-
tively). (The lines are the best fitting power relationships. 7=intertrial interval in seconds.)
tions, Fs(l, 18) =43.0 and 16.55, ps < .001, was a shift toward more control over sus-
for the 1st and 15th session, respectively. tained responding by the trial duration. Re-
The shortening of latency with continued sponse rates stabilized quite rapidly and
training is reflected here in a lower level showed no effect of absolute intertrial in-
and somewhat steeper slope for the function terval value and little effect of the ratio of
describing the 15th-session data. There is / to T, but substantial decrements at long
also a suggestion of a floor effect with la- T durations. Latency of the first response
tencies rarely exhibited below about 1 sec. of the trial was also insensitive to / values
Such an effect would result in some curva- or I/T ratios and showed a modest relation-
ture in the relationship for the ISth-session ship to T values.
data, but a lack-of-fit test performed on both The positive relationship between latency
functions failed to reveal reliable nonlinear- and T was the only indication that the dura-
ity. Thus, the power relationship appears to tion of the trial was "timed." However, the
be a reasonable description of the control failure to find an increase in response rate
exerted by trial duration on the latency of toward the end of the signal argues against
the first peck. timing in any accurate sense. Particularly
in view of the finding of depressed respond-
Discussion ing late in the trial at long T values, it is
doubtful that longer latencies with longer
The central finding of Experiment 1 was values of T reflect any sharp differentiation
that relative rather than absolute durations of when reinforcement is due.
of / and T controlled the speed with which Our previous report on the intertrial in-
autoshaped responding was acquired. The terval effect (Terrace et al., 1975) suggested
number of trials to acquisition was approxi- that rapid acquisition at long / values oc-
mately a power function of I/T. The ratio curred because the trial signal represented
relationship covered nearly two orders of a larger improvement in "predictiveness"
magnitude, with a suggestion that at very than when the intertrial interval was short.
large ratios acquisition is not as rapid as a One construction of predictiveness is ob-
strict power relationship would predict. Pos- tained by discretizing time in the session
sibly, very large ratios retard acquisition. A into subintervals containing trials and sub-
recent report (Griffin, 1975) describes ac- intervals containing intertrial intervals, and
quisition as a U-shaped function of I/T, correlating these with reinforcement and
with the descending portion a power func- nonreinforcement. This discretized time as-
tion comparable to the one we report. sumption, under appropriate restrictions,
Once autoshaped pecking emerged, there may be shown to result in increasing cor-
272 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
relation with increasing values of / (Gibbon, lected because it meant that most subjects
Berryman, & Thompson, 1974). However, would experience all of these intervals at
this sort of analysis ignores the potential least once before key pecking emerged. If
contribution of temporal cues to predictive- an internal representation of these variable
ness. From the strictly statistical point of intertrial times is somehow averaged by
view, the correlation between trial signal subjects before conditioning is complete,
and food remains the same whether inter- then the question arises as to what sort of
trial durations are variable or fixed, as long averaging (e.g., arithmetic, geometric) it is.
as they have the same mean values. An An additional aim of Experiment 2 was to
alternative view would regard temporal un- provide data on this point, since fixed /
certainty about trial onset as a potential con- values are not averaged.
tributor to learning the association between
the signal and food. In the extreme case, Method
if subjects were able to estimate very accu-
rately the time at which food delivery was Subjects. The subjects were 64 white Carneaux
scheduled, a trial signal indicating the same pigeons approximately 9 months old at the start
of training. Throughout the experiment, each sub-
thing would be redundant. Accordingly, ject was maintained at 80% ± 3% of its free-
blocking of conditioning to the trial signal feeding weight.
might be observed, similar to the well-docu- Apparatus. The apparatus was the same as that
mented blocking of redundant cues in fear used in Experiment 1.
conditioning (Kamin, 1969; Rescorla & Procedure. Subjects were assigned in groups
of four to four I and T values combined factorially
Wagner, 1972). (7 = 48, 96, 192, and 384 sec; T = %, 12, 16, and
A blocking hypothesis of this sort is spec- 32 sec). The factorial combinations resulted in
ulative. A demonstration would require pre- four I IT ratios (I/T = 3, 6, 12, and 24) that con-
training with a fixed interreinforcement tained more than one group. This provided a basis
for assessing whether the ratio relationship ob-
interval and subsequent assessment of re- served in Experiment 1 would hold for the fixed
tarded acquisition of autoshaping at that in- intertrial condition.
terval relative to control groups studied at The magazine-training and autoshaping-training
comparable testing parameters. However procedure in Experiment 2 was the same as that
followed in Experiment 1 except for the constant
indirect evidence on this point is provided duration of the intertrial interval.
by Experiment 2 reported below, in which
intertrial intervals were fixed at values
equal to the mean intertrial intervals of Ex- Results
periment 1. If temporal predictiveness is a Acquisition. Experiment 2 replicated the
feature of association learning in autoshap- finding of Experiment 1 that both increasing
ing, comparable to context effects (Tomie, the intertrial duration and decreasing the
1976), then temporal cues under fixed I trial duration resulted in more rapid acqui-
might compete with the key-light signal, re- sition. Acquisition scores from all birds (the
sulting in retarded acquisition. number of trials to satisfy the acquisition
criterion) were logged and an analysis of
Experiment 2 variance performed on the data. The inter-
trial interval and trial duration effects were
In Experiment 1, temporal cues for re- large, Fs(3, 44) = 7.36 and 7.95, ps < .001,
inforcement were reduced by making / for / and T, respectively, and no interac-
variable. In Experiment 2, we studied fixed tion was present. The acquisition data for
intertrial interval durations over a range of each bird are presented in Appendix C.
/ values spanning the center of the ratio A central finding for Experiment 1 was
relation (Figure 4). The variable intertrial that acquisition speed was approximately
times programmed in Experiment 1 were constant for constant values of the ratio of
scheduled to repeat the sampling distribu- the arithmetic mean of / to T. Figure 9
tion every 25 trials. This number was se- shows median trials to acquisition as a func-
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 273
in the signal is greatest at / = 384 sec and pronounced at the shorter 7 values and less
shows a more gradual approach to higher clear at the long 7 values.
rates than at the smaller 7 values. The effect Average response rates from the last 4
is roughly proportional; the reduction seems days of training were logged and subjected
to occur in the first one or two tenths of the to analysis of variance. There was a clear
signal independently of the absolute size of effect of T value, F(3, 46) = 3.93, p <
the signal (compare, for example, T = 16 .025, and no effect for 7 value or interac-
sec and T = 32 sec for 7 = 384 sec). tion. The control by trial duration may be
As in Experiment 1, responding on the observed in Figure 12, which shows on
1st day is quite similar to responding over double log coordinates the rate for each
the first 4 days, and these values are gen- group as a function of trial duration. The
erally somewhat lower than the rates sub- best fit power relation, represented by the
jects exhibit at the end of training. Again, line, accounts for about 38 % of the variance
also, the effects of trial duration are more in these data.
complex. There is a tendency toward re- Figure 13 presents group median latency
duced absolute levels of responding with in- from the 1st and 15th sessions as a function
creasing trial duration, and this finding is of trial duration. Both axes are logged, and
1 = 48
1=96
7
r 1=192
. ^^ 1=384
10 I 5 10 I 5 10 I 5 10 I
TENTHS OF THE TRIAL
Figure 11. Response rates on a log scale in successive tenths of the trial signal at each combi-
nation of intertrial (/) and trial (T) duration (in seconds). (/ values are constant within
rows and increase from top to bottom. T values increase from left to right. The average within
each / value is at the extreme right. Rates for the 1st day are indicated by Xs, for the first
4 days by open circles, and for the last 4 days by filled circles.)
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 275
l T
e 48 8
Ist SESSION 15th SESSION 48 12
I 48 16
A 48 32
10 •
96 8
I 96 12
0 96 16
4 96 32
o 192 8
o 192 12
D 192 16
oo 192 32
t 384 8
• 384 12
« 384 16
• 384 32
6 10 20 40 6 10 20 40
circles represent data from the first 4 days rates that all groups exhibit during the
after acquisition of responding in the trial, trial signal.
and filled circles are data from the final 4 A clear trend in the character of these
days of training. Functions are not presented functions may be discerned from the right-
unless rates were above .5 per min. This most column, which represents the average
was the case for 11 of the 16 groups after at each 7 value. When I is short, responding
prolonged training. However, all groups increases over successive fifths of the inter-
showed some responding early in acquisi- trial interval so that response rates are high-
tion. Of the five groups showing responding est just before the next trial signal is due
after prolonged training, only Groups / = to be presented. As / increases, however,
48 sec show responding at any substantial this trend diminishes and is replaced by the
level (about 4 responses per min). These reverse (concave upward) form in which
rates, while maintained throughout training, response rates decline monotonically as the
are about 10 to 20 times lower than the intertrial interval elapses. The two inter-
2
I
0.4
1=96
1=192
0.4
2
I 1=384
0.4
0.3
L
0.1
I 3 5 13 5 1 3 5 1 3 5 i 3 5
FIFTHS OF THE TRIAL
Figure 14. Response rate on a log scale in successive fifths of the intertrial interval (/) at each
combination of / and trial duration (T) for Experiment 2. (/ values are constant within a row
and increase from top to bottom. T values increase from left to right. The average within a
given / value is shown on the extreme right. Open circles are data from the first 4 days after
responding to the trial signal was initiated, and filled circles are data from the last 4 days of
training.)
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 277
mediate / values show intermediate function when it is small associative strength is ac-
forms, with 7 = 96 sec resulting in roughly quired slowly or not at all.
constant rates over the intertrial interval. An alternative view might hold that ac-
This change in the form of the functions ap- quisition scores reflect both associative
pears also to be representative of particular learning about the signal and performance
I, T combinations. variables which affect how strongly that as-
The change in curvature from concave sociative learning is indexed. For example,
down to concave up with increasing 7 values we will argue below that at very short trial
may result in part from pooling over a durations birds may be unable to show con-
larger absolute amount of postreinforcement ditioned responding when associative con-
time when calculating rates in successive ditioning might nevertheless be present.
fifths of the larger 7 values. For example, Differences across training conditions are
the / = 48 sec function corresponds in ab- then confounded with differences in the con-
solute time to the first half of the 7 = 96 sec ditions under which associative responding
function, which in turn corresponds to the is assessed. The present experiments (and
first half of the 7 = 192 sec function, and so acquisition studies in general) are not de-
on. Data were not collected in absolute time signed to discriminate these alternatives,
since reinforcement, and so the point may and thus it remains possible that key peck-
not be evaluated definitively, but a compari- ing, once acquired, reflects both the training
son of the first two points in the 1 — 96 sec conditions and factors unrelated to associa-
function with the first four points in the 7 tive conditioning that might influence per-
= 48 sec function suggests that at least the formance of the response. However, some
absolute levels of these functions are chang- features of the data argue against the latter
ing reliably with 7 value. Responding is interpretation. Sustained responding, both
about three or four times as great for early and late in training in both experi-
Groups 7 = 48 sec than for Groups 7 = 96 ments, was insensitive to the intertrial in-
sec, and curvature is evident throughout the terval variable. If nonassociative perform-
range for the smaller 7 value as well. There ance factors were involved in early condi-
is also a tendency, evident in the individual tioned responding, one might expect those
functions for different T values, for inter- same factors to show continued control over
trial responding to be lower with longer T sustained responding. However, continued
values, paralleling this same trend for re- training resulted in a modest degree of con-
sponding during the signal (Figure 11). trol over responding by the trial duration,
but virtually no control over responding by
the intertrial duration. Yet these two factors
General Discussion interact powerfully in acquisition. It seems
The central finding in both experiments more likely to us that the emergence of re-
was that relative rather than absolute trial sponding here reflects quite directly the
and intertrial times controlled the speed power of the training conditions to estab-
with which responding to the signal for re- lish an associative link between key light
inforcement emerged. One interpretation of and food. When associative strength is suf-
this result would regard the point in train- ficient, the signal functions as a "sign stim-
ing at which responding first occurs as an ulus" (Hearst & Jenkins, 1974) and re-
index of equivalent associative strengths for sponding is evoked.
the signal across different training conditions. The interaction of trial and intertrial du-
According to this view, pecking results ration in the acquisition of responding was
once a threshold of associative strength is evident over a range of nearly two orders
reached. However, this threshold is reached of magnitude. In this range, subjects seemed
at different rates under different training to require a number of reinforced trials be-
conditions. When the I/T ratio is large, fore pecking, which was approximately a
associative strength is acquired rapidly, and power function of the I/T ratio. This rela-
278 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
tionship held both when the intertrial inter- Quite apart, however, from the form of
vals were variable and / values were the the relationship between N and I/T, the
arithmetic mean of the variable sample and ratio effect itself places a powerful con-
when intertrial intervals were fixed. At val- straint on any theory which purports to
ues below //T—1.0, birds did not acquire encompass this result. The ratio effect is
at all, and at I/T — 96, acquisition appeared reminiscent of Weber's law for time dis-
to be retarded relative to the power func- crimination (e.g., Stubbs, 1968) in that the
tion prediction. The reasons for these limits trial duration must represent a constant
are probably not the same. At the lower fraction of the intertrial duration for com-
limit, when trial durations are as long or parable speeds of conditioning. However,
longer than intertrial periods, subjects may temporal discrimination in the sense of ac-
never learn the association between signal curate timing does not appear to be in-
and food for the same reasons that this volved. Two considerations force that con-
learning is retarded as the ratio of 7 to T clusion. First, accurate timing of the trial
is reduced. That is, the degradation of asso- duration might be expected to result in more
ciative value with decreasing I/T may sim- rapid responding toward the end of the sig-
ply exceed the threshold for autoshaped key nal, when reinforcement is due. This does
pecking at ratios of about 1.0. not develop to any substantial degree until
At the upper end, the ratio relationship late in training—and then only when / val-
may fail for several different reasons. Large ues are long and fixed. Early in training,
ratios perforce require either very short responding—once it has begun in the trial—
trial values or very long intertrial values. appears to be roughly constant throughout
If T is less than or equal to 2 sec, birds may the remainder of the signal. Since only the
not be functionally in a position to deliver acquisition data are controlled by the trial-
a response with a short enough latency for to-intertrial ratio, some alternative construc-
it to be recorded during a trial, particularly tion is required to explain this result.3
when trial onset times are variable, as in Additional evidence is provided by the
Experiment 1. Thus, subjects may acquire acquisition function of Experiment 2. If
some associative strength to the signal yet some internal representation of time values
may not be able to index that strength with of the trial relative to the intertrial time are
our dependent variable measure. Conversely, important in acquisition, then it is surely
when trial durations are moderate and inter- not an accurate time representation or we
trial durations are very long, subjects may would have observed some effect when the
not be "attending" when the trial signals intertrial interval was constant. If the num-
come on, due to roosting or other activities ber of trials before acquisition occurs were
in the experimental chamber. This may re- sufficiently large, one might expect some
sult in functionally fewer reinforced trials, temporal conditioning to the constant inter-
thus again retarding acquisition. reinforcement interval in the fixed / case,
In the center portion of the I/T range, but not to the variable interreinforcement
the speed of acquisition is reasonably well intervals programmed in Experiment 1.
described by a power function. We do not Since the acquisition functions from the two
presently know why this relationship, as experiments were indistinguishable, it seems
opposed to some other monotone decreasing that the number of trials before responding
function of I/T, is the proper one. Regres- begins is insufficient to substantially reduce
sion analysis of the log of the number of
trials to acquisition as a function of un- 3
Data from a companion experiment with prob-
logged ratio values results in considerably abilistic reinforcement (Farrell, Terrace, Gibbon,
more scatter and lower correlations. Of & Locurto, 1974) suggest that timing of the trial
course, this simply means that a simple ex- signal develops more powerfully with extended
ponential relationship to I/T is less appro- training under partial reinforcement. However,
these data confirm the present finding that no time-
priate than the power relationship that we correlated changes in response rate are observed
exhibit. immediately after pecking emerges.
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 279
variability in the internal representation of there was some tendency for responding to
the intertrial interval when that interval is occur after reinforcement, followed by a de-
constant. Thus, whatever mechanism is pro- cline as the intertrial interval elapsed. If the
posed for specifying the occurrence of the increase observed at short 7 values reflects
key peck, it cannot depend critically on ac- anticipation, then the decrease as 7 elapses
curate timing of the two intervals. at long 7 values may reflect inability to an-
The independence of acquisition from the ticipate trial onset. In neither case—an-
regularity or irregularity of trial spacing is ticipation or failure to anticipate—did the
further reflected in comparable variability of acquisition of responding reflect these differ-
acquistion scores in the two experiments. ences. Rather, intertrial responding devel-
Evidently, birds experiencing variable inter- oped only after trial responding was es-
trial periods integrate them in a linear man- tablished.
ner so that their effect is comparable to Latency of the first trial response in the
that on birds receiving a single 7 value two experiments differed somewhat in its
equal to the arithmetic mean of the variable dependence on T. In Experiment 1, in which
sample. If internal representations of trial trial onset was variable and thus unantici-
and intertrial times are averaged over suc- pated, latency depended more on the dura-
cessive experiences when / is variable, then tion of the trial. In Experiment 2, this
the similar acquisition in the two experi- dependence was attenuated, particularly at
ments argues against a logarithmic repre- relatively short T values. Such a finding
sentation of time (a "Fechnerian" scale in may reflect an increase in readiness to re-
the sense of Luce & Galanter, 1963). An spond relatively early in the trial when the
acquisition function based on the geometric trial onset is—however imprecisely—antici-
mean of our sample (Appendix A) would pated. Why a dependence on T should be
be off by a factor of about two from the evident at all, however, is not clear. Cer-
acquisition function we obtained in Experi- tainly, latency here does not reflect timing
ment 2. The power function representation of the duration of the signal, since the la-
would not be altered by a geometric averag- tency distributions are invariably highly
ing procedure, since our variable samples skewed toward the low end of the trial and
were computed as multiples of a standard were "complete" well before the trial had
sample. Thus, a Fechnerian internal repre- elapsed. Also, response rates over successive
sentation of the relevant time values might portions of the trial failed to indicate any
well lead to the same power function as a clear evidence of timing.
linear internal representation of the relevant The main results of these experiments
times values. What would differ, however, bear importantly on theoretical positions
is the level of the two functions. A linear that would tie the autoshaping phenomenon
internal representation should result in the to stimulus-stimulus contingency learning
comparability between Experiment 1 and in general. Below we sketch two theoretical
Experiment 2 that we observe, while a loga- approaches and the constraints that the pres-
rithmic transform should alter the levels of ent data force upon them.
the functions for the two experiments in the
direction of more rapid acquisition in the
fixed 7 case. Re scoria-Wagner Theory
The constant intertrial interval used in Rescorla and Wagner (1972) and Wag-
Experiment 2 resulted in very similar pat- ner and Rescorla (1972) have proposed a
terns of acquisition, but some differences in theoretical account of association learning
maintained behavior. At the shortest 7 value which may be adapted to the autoshaping
(48 sec), some anticipation of trial onset situation. In their theory, excitatory control
was evident in the pattern of increased re- is acquired by a conditioned stimulus (CS)
sponding over the intertrial period. At the which is reinforced in compound with back-
longer 7 values, intertrial responding was ground stimuli. The speed with which the
not maintained with prolonged training, but stimulus acquires its associative value is
280 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
directly related to two stimulus parameters: no good reason why they must counter-
a learning rate parameter that is character- balance in just equal measure.
istic of the subject and the reinforcer and a The finding of some dependence between
salience parameter that roughly interprets sustained responding and trial duration is
as the "conspicuousness" of the CS. The consonant with Rescorla-Wagner theory if
background stimuli also increase in associa- one regards the discretized time unit as a
tive value, at a rate determined by their reasonable manner for estimating differen-
salience, as they are reinforced in com- tial reinforcement probabilities. On this con-
pound with the CS. The central feature of struction, long trials result in low probabili-
the theory is that growth of the compound ties of reinforcement and in the theory are
is proportional to the discrepancy between associated therefore with low asymptotes of
the sum of the current associative strengths associative value. On the other hand, the
of all stimuli present and the obtainable dependence between rate of acquisition and
asymptote. This results in an exponential T value is not very strong, while depen-
growth function. In this context, the effect dence between asymptote and T value in
of the intertrial interval is to introduce un- the theory is direct.
reinforced "trials" on the background stim-
uli alone during the time intervening be-
tween reinforced background-plus-CS trials. Contingency Theory
Qualitatively, it is clear that increasing the A second view of associative learning that
amount of extinction on the background re- is somewhat less precise in its detail evolves
sults in a lower compound associative value from Rescorla's (1967, 1968) analysis of
of trial plus background on a subsequent the role of correlation in establishing ex-
reinforced trial. The result is that these citatory/inhibitory fear conditioning. Gibbon
trials increase associative value relatively et al. (1974) proposed a quantitative metric
more rapidly (per reinforcement) than for the correlational view summarized in
trials in which the background has not been root mean-square contingency, <£. This co-
extinguished to such low levels. In the efficient is argued to reflect the degree of
Rescorla-Wagner theory, temporal param- potential association that a given paradigm
eters therefore influence the rate at which may engender. A discretized time assump-
associations are learned but not the eventual tion is again required here to dichotomize
level of associative strength. The present conditioning sessions into trial units of CS
acquisition data are consonant with this and no CS, and cross-classify these events
view in that long 7 values facilitate acquisi- against reinforcement and nonreinforcement.
tion. The main features of this correlational
The ratio effect, however, strains this ac- analysis are consonant with current data
count. If we discretize time into "trial on contingency learning in autoshaping
units," increasing the trial duration corre- (Gibbon, Locurto, & Terrace, 1975), and
sponds to decreasing the probability of re- several features of the present data appear
inforcement, while increasing the intertrial compatible with this approach also. Increas-
interval corresponds to providing more fre- ing / corresponds to increasing the cor-
quent extinction trials for the background relation between signal and food. More im-
stimuli. It is readily shown that decreasing portantly, if reinforcement probability is
probability of reinforcement lowers the maintained constant, the correlation between
asymptote as well as the rate of approach, signal and reinforcement is unchanged when
and increasing the frequency of extinction the intertrial-to-trial durations are main-
trials by increasing I increases the effec- tained in the same ratio. While at first blush
tiveness of each reinforcer. Thus, one might this would seem to handle the I/T ratio
argue that the effects of lowered probability effect, on reflection, the discretized time-
of reinforcement are counterbalanced by the unit assumption may be seen to violate an-
effect of increasing /. However, there seems other feature of the correlation analysis.
TRIAL AND INTERTRIAL DURATIONS IN AUTOSHAPING 281
When T and / periods are both extended Ergun, S. Application of the principle of least
and the discretized time unit is maintained squares to families of straight lines. Journal of
Industrial and Engineering Chemistry, 1956, 48,
constant, not only are trial and intertrial 2063-2068.
frequencies increased, but reinforcement Farrell, L., Terrace, H. S., Gibbon, J., & Locurto,
probability is decreased. Unless reinforce- C. Partial reinforcement in auto-shaping: Acqui-
ment rate per trial unit is maintained con- sition, maintenance and extinction. Proceedings
stant while trial and intertrial times are of the 45th Annual Convention of the Eastern
Psychological Association, 1974, 74, 183.
changed, effects may be expected in the di- Gibbon, J., Berryman, R., & Thompson, R. L.
rection of lowered correlation values with Contingency spaces and measures in classical
increased absolute trial duration. Thus, and instrumental conditioning. Journal of the
while in a general way the correlational ap- Experimental Analysis of Behavior, 1974, 49,
732-764.
proach seems appropriate, the strong con- Gibbon, J., Locurto, C., & Terrace, H. S. Signal-
straint of the ratio effect appears to com- food contingency and signal frequency in a con-
promise this view as well. tinuous trials auto-shaping paradigm. Animal
While these theoretical positions are some- Learning and Behavior, 1975, 3, 317-324.
Gormezano, L, & Moore, J. W. Classical condi-
what strained by the ratio effect, neither tioning. In M. Marx (Ed.), Learning: Pro-
are embarrassed by the similiarity of the cesses. London: Macmillan, 1969.
functions obtained in Experiments 1 and 2. Griffin, R. W. Classical conditioning of directed
That is, fixed or variable / values, on the motor behavior in rats and pigeons: Effects of
intertrial interval duration and variation. Pro-
discretized time assumption, make no dif- ceedings of the 46th Annual Meeting of the
ference to correlation calculations, nor to Eastern Psychological Association, 1975, 75, 79.
learning rates according to the Rescorla- Hearst, E., & Jenkins, H. M. Sign-tracking: The
Wagner theory. Both of these accounts are stimulus-rcinforcer relation and directed action
(Monograph of the Psychonomic Society). Aus-
insensitive to temporal regularities and so tin, Tex.: Psychonomic Journals, 1974.
the results of the two experiments are con- Kamin, L. J. Predictability, surprise, attention and
sonant with either position. conditioning. In B. A. Campbell & R. M. Church
The ratio effect may require some funda- (Eds.), Punishment and Aversive Behavior,
New York: Appleton-Century-Crofts, 1969.
mental rethinking about the associative Luce, R. D., & Galanter, E. Discrimination. In
learning process in autoshaping. At the very R. D. Luce, R. R. Bush, & E. Galanter, Hand-
least, the translation of our theoretical con- book of mathematical psychology (Vol. 1).
structs in time requires elaboration. One ap- New York: Wiley, 1963.
proach might regard the subjective trial Pearson, E. S., & Hartley, H. O. (Eds.) Bio-
metrika tables for statisticians (Vol. 1). Cam-
unit as rescaled when trial durations are bridge: Cambridge University Press, 1970.
changed. Subjects must then be regarded Perkins, C. C., Jr., Beavers, W. O., Hancock,
as calculating the frequency of unreinforced R. A., Jr., Hemmendinger, P. C., Hemmen-
intertrial periods by a comparison with the dinger, D., & Ricci, J. A. Some variables affect-
ing rate of key pecking during response-indepen-
trial duration. While an elaboration of this dent procedures (autoshaping). Journal of the
view may eventually turn out to be correct, Experimental Analysis of Behavior, 1975, 24,
at this level of analysis, it does little more 59-72.
than rename our ratio effect. We need to Rescorla, R. A. Pavlovian conditioning and its
proper control procedures. Psychological Re-
understand how the scaling process works view, 1967, 74, 71-80.
and how the comparison between trial and Rescorla, R. A. Probability of shock in the pres-
intertrial, or trial and background, is made. ence and absence of CS in fear conditioning.
Journal of Comparative and Physiological Psy-
chology, 1968, 66, 1-5.
Rescorla, R. A., & Wagner, A. R. A theory of
References Pavlovian conditioning: Variations in the ef-
fectiveness of reinforcement and non-reinforce-
Brown, P., & Jenkins, H. M. Autoshaping of the ment. In A. Black & W. Prokasy (Eds.), Clas-
pigeon's keypeck. Journal of the Experimental sical conditioning. New York: Appleton-Cen-
Analysis of Behavior, 1968, 11, 1-8. tury-Crofts, 1972.
Draper, N. R., & Smith, H. Applied regression Staddon, J. E. R., & Simmelhag, V. L. The "su-
analysis. New York: Wiley, 1966. perstition" experiment: A re-examination of its
282 GIBBON, BALDOCK, LOCURTO, GOLD, AND TERRACE
implications for the principles of adaptive be- context conditioning. Journal of Experimental
havior. Psychological Review, 1971, 78, 3-43. Psychology: Animal Behavior Processes, 1976,
Stubbs, A. The discrimination of stimulus dura- 2, 323-334.
tion by pigeons. Journal of the Experimental Wagner, A. R., & Rescorla, R. A. Inhibition in
Analysis of Behavior, 1968, 11, 223-238. Pavlovian conditioning: Application of a theory.
Terrace, H. S., Gibbon, J., Farrell, L., & Baldock, In R. A. Boakes & M. S. Halliday (Eds.), In-
M. D. Temporal factors influencing the acquisi- hibition and learning. New York: Academic
tion and maintenance of an autoshaped keypeck. Press, 1972.
Animal Learning and Behavior, 1975, 3, 53-62. Winer, B. J. Statistical principles in experimental
Tomie, A. Interference with autoshaping by prior design (2nd ed.). New York: McGraw-Hill, 1971.
Appendix A
Table Al
Intertrial Interval (I; in sec) Comprising
the Sample for I = 96
Appendix B
Table Bl
Acquisition Data for Each Subject: Experiment 1
Table Bl (continued)
Number of trials to Number of trials to
1st 1st satisfy 1st 1st satisfy
response response acquisition response response acquisition
1 T I/T Bird in trial in / criterion / T I/T Bird in trial in / criterion
48 4 12 Y.422 16 18 19 96 32 3 Y.319 61 64 65
Y.424 10 14 13 Y.409 90 85 104
Y.426 15 17 18 Y.414 17 6 21
Y.429 7 11 10 Y.416 51 55 54
Y.741 24 27 26 Y.677 151 144 163
Y.742 32 37 34 Y.695 69 76 71
Y.743 5 9 7 Y.705 53 62 67
Y.744 17 29 23 Y.708 33 42 40
Appendix C
Table Cl
Acquisition Data for Each Subject: Experiment 2
Number of Number of
trials to trials to
satisfy satisfy
acquisition acquisition
I T I/T Bird criterion I T I/T Bird criterion