Can existing associative principles explain occasion setting? Some old ideas and some new data

Abstract

Since occasion setting was identified as a type of learning independent of 'simple' associative processes, a great deal of research has explored how occasion setters are established and operate. Initial theories suggested that they exert hierarchical control over a target CS→US association, facilitating the ease with which a CS can activate the US representation and elicit the CR. Later approaches proposed that occasion setting arises from an association between a configural cue, formed from the conjunction of the occasion setter and CS, and the US. The former solution requires the associative principles dictating how stimuli interact to be modified, while the latter does not. The history of this theoretical distinction, and evidence relating to it, will be briefly reviewed and some novel data presented. In summary, although the contribution of configural processes to learning phenomena is not in doubt, configural theories must make many assumptions to accommodate the existing data, and there are certain classes of evidence that they are logically unable to explain. Our contention is therefore that some kind of hierarchical process is required to explain occasion-setting effects.

Full text

Can existing associative principles explain occasion setting? Some old ideas and some
new data
Charlotte Bonardi1, Jasper Robinson1 & Dómhnall Jennings2
1 School of Psychology, University of Nottingham
2 Institute of Neuroscience, Newcastle University
Since occasion setting was identified as a type of learning independent of 'simple'
associative processes, a great deal of research has explored how occasion setters are
established and operate. Initial theories suggested that they exert hierarchical control
over a target CS→US association, facilitating the ease with which a CS can activate the
US representation and elicit the CR. Later approaches proposed that occasion setting
arises from an association between a configural cue, formed from the conjunction of the
occasion setter and CS, and the US. The former solution requires the associative
principles dictating how stimuli interact to be modified, while the latter does not. The
history of this theoretical distinction, and evidence relating to it, will be briefly reviewed
and some novel data presented. In summary, although the contribution of configural
processes to learning phenomena is not in doubt, configural theories must make many
assumptions to accommodate the existing data, and there are certain classes of evidence
that they are logically unable to explain. Our contention is therefore that some kind of
hierarchical process is required to explain occasion-setting effects.
Keywords: occasion setting; hierarchical theory; configural theory; learning.
1

1. Introduction
Skinner (1938) was the first to suggest that performance of an operant response
could come under the control of a discriminative stimulus. The idea spread to the
cognitive behaviourist literature, with Holland (e.g. 1983; 1985) and Rescorla (e.g.
1985; 1986) demonstrating the same effect in Pavlovian conditioning. Specifically, after
training that a conditioned stimulus (CS) signalled an unconditioned stimulus (US),
animals could confine performance of their conditioned response (CR) to the CS in the
presence of a stimulus termed a modulator (Rescorla, 1985), a remote initiating stimulus
(Jenkins, 1985) or an occasion setter (OS; Holland, 1983).
The truly novel finding was that the control exerted by the OS was independent
of its associative properties. An OS could enhance performance of a CR that it could not
itself elicit (Ross & Holland, 1981; Rescorla, 1985), even after OS extinction (Holland
1989a), and pretraining the OS to signal the US impeded its ability to acquire control
over the CR (e.g. Rescorla, 1986)1. This suggested that an OS's effect on behaviour did
not rely on activation of the US, CS or CR - that it was not mediated by standard
associative principles2. Two classes of theory emerged in response to this challenge to
associative theory, which had until this point dominated accounts of learned behaviour.
The first assumed additional, nonassociative principles must be invoked, giving rise to
the US modulation, memory systems and hierarchical accounts. The second asserted that
existing associative principles could explain occasion setting - provided combinations of
stimuli could be represented and be subject to associative learning; this class includes
the various versions of configural theory (e.g. Rescorla, 1972; Brandon, Vogel &
Wagner, 2000; Pearce, 1987; 1994). As independent evidence for such configural
1 probably because such pretraining blocked acquisition of associative strength by the
target CS (see Swartzentruber, 1995 for a review of related findings)
2 This should not be taken to imply that an OS may not also have associative properties
that influence behaviour, simply that its action cannot be explained solely in those terms.
2

theories accumulated (e.g., Haselgrove et al., 2008; Pearce, et al. 2002; Williams,
Sagness & McPhee 1994), this seemed the more parsimonious explanation, with the
result that alternative accounts of occasion setting were eclipsed. This article will revisit
evidence relating to these issues, and evaluate the extent to which associative theory can
explain occasion setting, or whether additional nonassociative principles are required.
We begin with the US Modulation and Hierarchical Accounts theories of
occasion setting, for which the evidence, predominantly favouring the hierarchical
account, is described. The Configural Learning alternative to the hierarchical approach,
and its failure to account for evidence of US and CS/US specificity, is then considered;
after this the evidence on Mechanisms of Occasion-Setter Formation which challenged
the hierarchical approach is reviewed. Elaborations of Hierarchical and Configural
Theories are then presented that can, with added assumptions, explain most of the
existing data. Some Further Discriminating Evidence that could allow us to choose
between these elaborated theories is discussed. We conclude by considering whether or
not there is a need to suppose hierarchical processes to explain occasion-setting effects.
2. US modulation and hierarchical accounts
The first key theories were the US modulation (Rescorla, 1985) and hierarchical
accounts (Holland, 1983). US modulation elaborated on the existing conceptualisation of
a conditioned inhibitor (a stimulus predicting the omission of an otherwise expected US,
and counteracting the effect of CSs predicting that US; e.g. Konorski, 1948; Rescorla,
1969) as acting through suppression of activation in the US representation. The US
modulation account proposed the complementary process, that a positive occasion setter
lowers the activation threshold of the US representation, increasing its sensitivity to
excitatory cues. This allows the CS to activate the US representation with greater ease in
3

the presence of the OS than in its absence. In contrast, the hierarchical account asserted
that the OS facilitates operation of the association between CS and US (Figure 13).
Figure 1 about here
The accounts may be discriminated in terms of transfer. Suppose an OS signals
that cs1 predicts US1 (Figure 1). If the OS facilitates activation of US1, it will enhance
responding to cs2 associated with US1, but be without effect on cs3 associated with US2 -
it is US-specific, not CS-specific. But if the OS enhances operation of the cs1→US1
association it will have no effect on cs2, even if cs2 is also associated with US1: it will be
both CS-specific and US-specific. Evidence suggests that occasion setters are both CS-
and US-specific, and that the extent to which such specificity is observed is influenced
by procedural factors (Swartzentruber, 1995).
2.1 CS-specificity Many studies have shown that OS1 signalling that cs1 predicts US1
(OS1: cs1→US1), may control the CR to a cs2 that also predicts US1. But this transfer is
typically incomplete: OS1 is rarely as effective with cs2 as with cs1 (e.g. Davidson &
RescorIa, 1986; Holland, 1986, 1989b, 1989c; RescorIa, 1985; see Swartzentruber,
1995, for a review). This is inconsistent with US modulation: if OS1 facilitates activation
of US1 it should modulate all CSs associated with US1 equally. But if cs2 were to suffer
generalisation decrement through being combined with OS1, this could reduce
responding to cs2, allowing US modulation to explain the incomplete transfer. This
suggestion has not survived experimental test, however: Bonardi (1996) trained pigeons
3 An alternative conceptualisation of the hierarchical account is that as an occasion-set
CS is typically both reinforced and nonreinforced during training, it must have both
excitatory and inhibitory associations with the US, and that the occasion setter inhibits
the inhibitory association (Bouton & Nelson, 1998). However this account assumes that
occasion setting is impossible if the occasion-set CS has no inhibitory strength, and there
is evidence against this position (e.g. de Brugada et al., 1995; Hall & Honey, 1989).
4

that two occasion setters, OS1 and OS2 (a tone and flashing houselight) signalled
reinforcement of keylights cs1 and cs2, respectively; cs1 and cs2 were nonreinforced when
presented alone. In contrast cs3 and cs4 were reinforced regardless of whether they were
accompanied by OS1 and OS2 or not (Table 1). Thus OS1 and OS2 were occasion setters
for cs1 and cs2, but not for cs3 and cs4. Then responding to cs1, cs2, cs3 and cs4 was
examined in the presence of OS1 and OS2, in combinations that were the same or
different from those of training. Incomplete transfer of occasion setting - more
responding on same (OS1:cs1, OS2:cs2) than on different (OS1:cs2, OS2:cs1) trials - was
observed. If this were due to generalisation decrement of cs1 when it was first presented
with OS2 (and of cs2 with OS1) then the same effect would be expected on trials with cs3
and cs4 - more responding on same (OS1:cs3, OS2:cs4) than different (OS1:cs4, OS2:cs3)
trials. In fact numerically the opposite was observed4 (Figure 2). If incomplete transfer
of occasion setting is not due to generalisation decrement, it implies that occasion setters
are CS-specific, contrary to the US-modulation account (cf. Rescorla, 1991a; 1991b).
Table 1 about here
Figure 2 about here
2.2 Memory systems This line of work raised further issues: Holland (1989c) reported
that OS1 signalling that cs1 predicts US1 (OS1: cs1→US1) modulated responding to cs2
only if it had been trained in an occasion-setting task. Conversely, OS1 modulated
responding to a stimulus that signalled US2, but only if US2 had been involved in an
occasion-setting task. Holland (e.g. 1989c) thus proposed the multiple memory systems
account: (i) events involved in occasion-setting are represented in a different memory
system from those that have not, (ii) transfer is more likely between events within a
4 This may be explained in terms of SOP (Wagner, 1981) as retrieval-generated priming
of cs3 and cs4 on same trials reducing their ability to elicit the CR - cf., Honey, Hall &
Bonardi, 1993.
5

memory system. However, many have reported substantial or complete transfer to CSs
that have not been targets of occasion-setting, creating a problem for this theory (cf.,
Swartzentruber, 1995).
2.3 Role of stimulus generalisation The literature on transfer reveals great variability in
whether transfer of an occasion setter to a different CS is obtained, which is only partly
attributable to the transfer CS's training history. Although transfer is often better to CSs
that have also been targets of occasion setting, this may be understood in terms of
simpler principles such as stimulus generalisation. For example, hierarchical theory
assumes OS1, signalling cs1 predicts US1, acts directly on the cs1→US1 link, and so
should have no effect on cs2→US1. But if cs1 and cs2 comprise unique and common
elements (i.e. cs1 = csacsc and cs2 = csbcsc), then both cs1→US1 and cs2→US1 share a
csc→US1 component, and OS1 could influence responding to cs2 via its effect on this
csc→US1 association. That transfer is found more often in pigeons with visual keylight
CSs than in rats with audio-visual cues (Swartzentruber, 1995) lends credence to this
view, to the extent that stimulus generalisation is likely to be greater within a stimulus
modality than between modalities. Such effects could also explain why transfer occurs
more readily to occasion-set CSs. Generalisation may occur between occasion setters:
OS1 acting on cs1→US1 may transfer more effectively to a cs2 that has been occasion-set
by OS2 because of generalisation between OS1 and OS2; if cs2 has not been occasion-set
this source of generalisation is not available. Also generalisation between CSs may not
be based solely on physical similarity, but also via their common training history -
acquired equivalence (Honey & Hall, 1989). For example, Bonardi and Hall (1994a)
examined generalisation from occasion-set cs1 to two further cues cs2 and cs3, where cs2
had also been the target of occasion setting, but cs3 had not (Table 2)5. After pairing cs1
5 cs3 was reinforced and extinguished (Table 2), to give it a similar training history to cs2
without endowing it with occasion-setting properties.
6

with food they found greater generalisation of conditioned responding to cs2 than to cs3
(Figure 3). They argued that the common training history of the occasion-set cues
increased their similarity, which fostered selective transfer between them.
Table 2 about here
Figure 3 about here
2.4 Summary Transfer of occasion setting across CSs occurs, but it is typically
incomplete. Transfer can be more substantial when the transfer CS had been occasion-
set, but also occurs when it has not. These findings support the hierarchical account,
which predicts occasion setters should be CS-specific, but that some degree of transfer
can occur via stimulus generalisation, either between training and transfer CSs or
between occasion setters, and based on stimulus similarity or stimulus training history.
These findings are, however, also consistent with configural theory to which we turn
next.
3. Configural theory
Configural theories were developed to explain performance on nonlinear
discriminations such as negative patterning, A→US / B→US / AB→nothing. According
to the standard associative assumption about summation, a compound stimulus is
equivalent to the sum of its parts, so presenting A and B together sums their associative
strengths. If CR is monotonically related to associative strength, accurate performance is
impossible because responding to AB must be higher than to A or B alone. Configural
theories abandon this summation assumption, but differ in how they conceptualise the
7

stimulus compound. For example, Rescorla (1972) proposed that AB comprises the
elements of A and B plus a third, configural cue, x that is only present when A and B co-
occur; thus negative patterning becomes A→US / B→US / ABx→nothing, meaning x can
acquire inhibitory strength and allow solution of the task. In contrast Pearce (1987;
1994) proposed that although AB is distinct from A and B, generalisation can occur
between them based on the proportion of elements they share; some have argued this
effectively conceptualises the AB compound as a subset of the elements of A and B
(Brandon, Vogel & Wagner, 2000). Brandon et al. (2000) combined these ideas in the
replaced elements model, according to which some elements of A and B are replaced by
elements unique to AB. All three can explain negative patterning, because all predict that
the total associative strength accrued to A and B is not the only source of responding on
AB trials.
Configural theory can thus explain many of the facts about occasion setting by
recasting an OS1: cs1→US1 / cs1→nothing discrimination as conditioning of a configural
cue P created by co-occurrence of OS1 and cs1. If what has been conditioned is not OS1
but P, manipulations of OS1's associative strength will leave the associative strength of
P relatively intact. Configural accounts can also explain CS specificity effects: in the
OS1: cs1→US1 / cs1→nothing discrimination, for example, the configural cue P is
reinforced, and transfer of OS1 to cs2, in an OS1: cs2 compound occurs to the extent that
there is generalisation between P and a second configural cue Q produced by co-
occurrence of OS1 and cs2. Many configural theories predict this would depend in part on
the similarity of cs1 and cs2 — which, as we saw above, is what seems to be the case.
3.1 US-specificity Both hierarchical and configural theories can thus explain occasion
setting and its CS-specificity. But they make different predictions about US-specificity -
whether OS1 trained as a signal that cs1 predicts US1 (OS1: cs1→US1) will act more
effectively on other CSs that also predict US1. The hierarchical account predicts that
8

occasion setters should be US-specific - yet Holland reported no sign of US specificity
(perfect transfer) with a CS that predicted US2, provided US2 had also been involved in
an occasion-setting task (Holland, 1989c). However, US specificity has been reported
even when US2 has been occasion-set: Bonardi, Bartle & Jennings (2012) trained rats on
two positive-patterning tasks in which OS1 signalled reinforcement of cs1 with US1, and
OS2 of cs2 with US2 (with cs1, cs2, OS1 and OS2 also presented alone - Table 3) and then
examined the ability of OS1 and OS2 to transfer to cs3 and cs4, which predicted US1 and
US2 respectively. Transfer was greater when the outcomes of the occasion setter and
transfer CSs were the same, despite the fact that both US1 and US2 had been involved in
occasion-setting tasks (Figure 4; see also Morell & Davidson, 2002).
Table 3 about here
Figure 4 about here
These findings are not consistent with the predictions of standard configural
theory which assumes that configural cues do not encode information about the USs that
they signal. This means they cannot explain how OS1, signalling cs1→US1, can transfer
more effectively to a stimulus that also signals US1 than to a stimulus that signals US2
(Morell & Davidson, 2002; Bonardi et al., 2012). In both cases responding should be
determined by the similarity of the trained configural cue OS1:cs1 to the OS1:cs2 cue
present at test, neither of which is affected by the nature of the US. Given the theoretical
importance of this issue, we conducted a further study in which 16 rats were trained on
two feature-negative discriminations with different USs (Table 4). OS1 and OS2 were
visual (illumination of two 2.8-W jewel lights pulsed at 1 Hz, or of a 2.8-W bulb
mounted inside the food magazine), cs1 and cs2 were auditory (white noise or 10-Hz
clicker both at 73 dB) and US1 and US2 were either 2 sucrose pellets or .3 ml of
9

groundnut oil (for complete description of apparatus see Bonardi et al., 2012). All
stimulus presentations were 10-s in duration, and there was a 5-s trace interval between
OS offset and CS onset on compound trials; the intertrial interval (ITI) comprised a fixed
60s plus a variable interval with a mean of 30s, and on reinforced trials US delivery was
delivered at CS offset. Each of the first 15 training sessions comprised, in a semi-random
order, 8 OS1→nothing, 8 OS2→nothing, 8 cs1→US1, 8 cs2→US1, 16 OS1: cs1→nothing
and 16 OS2: cs2→nothing trials6. In the next 18 sessions the OSs were also reinforced;
each of these sessions comprised 4 OS1→US1, 4 OS2→US1, 4 cs1→US1, 4 cs2→US2
trials, 24 OS1: cs1→nothing and 24 OS2: cs2→nothing trials. The final 6 of these sessions
also included three trials with each of two test excitors cs3 and cs4, a 300-Hz buzzer and
a 2-kHz tone, both at 75dB, paired with oil and sucrose respectively. These sessions
were otherwise identical to those of the preceding stage. The results from training are
shown in Figure 5. The rats learned the task, responding more to cs1 and cs2 when they
were presented alone on reinforced trials than on nonreinforced trials when they were
preceded by the OSs; ANOVA with trial type (CS reinforced or not) and session block as
factors revealed main effects of trial type, F(1, 15) = 6.67, MSe = 16.38, p = .021, block,
F(10, 150) = 16.36, MSe = 8.86, p < .001, and a significant interaction, F(10, 150) =
3.78, MSe = 1.82, p <. 001. Simple main effects analysis (using the pooled error term)
revealed an effect of trial type on blocks 3, 7, 10 and 11, smallest F(1, 15) = 4.62, MSe
= 16.38, p = .048. Finally, the mean rate of responding to cs3 and cs4 was 11.25 and
12.90 rpm for blocks 10 and 11 respectively.
Table 4 about here
6 For half the animals the noise was reinforced with sucrose and the click with oil and
for the remainder the reverse; for half of each of these subgroups the noise was preceded
by the jewel light, and the click by the tray light on compound trials, and for the
remainder the reverse.
10

Figure 5 about here
The ability of OS1 and OS2 to suppress responding to cs3 and cs4 was then
evaluated: in four test sessions rats continued to receive three trials with each test
excitor, plus 24 same trials in which either OS1 preceded cs3, or OS2 preceded cs4, and 24
different trials on which these combinations were reversed (12 of each type). If OS1's
effects were specific to US1 then OS1 should be more effective on same trials with cs3
that also predicted US1, than on different trials with cs4 which predicted US2. The test
session data are shown in Figure 6 in two, 2-session blocks; ANOVA performed on data
from the OS trials with trial type (same or different) and block as factors revealed a main
effect of trial type, F(1, 15) = 5.77, MSE = 2.29, p = .03, no effect of block, F < 1, and a
significant interaction, F(2, 30) = 4.70, MSE = .459, p = .047. Simple main effects
analysis revealed an effect of trial type on block 2, F(1, 15) = 5.72, MSE = 2.24, p = .03,
although not on block 1, F(1, 15) = 1.00, MSE = 2.24, p = .33. Thus, just as in the
positive patterning task, transfer was greater on same trials when the outcomes of the
occasion setter and transfer CSs matched, despite the fact that both US1 and US2 had
been involved in occasion-setting discriminations. Such US specificity cannot be
predicted by the configural account, but is perfectly consistent with hierarchical theory.
Figure 6 about here
3.2 Specificity to CS/US combination Another key discriminator between the
hierarchical and configural accounts is whether an OS1 which has signalled cs1→US1 and
cs2→US2 will be equally effective on cs1 if it is subsequently paired with US2, and on cs2
if it is paired with US1. If the occasion setter acts on the actual association between two
specific events, as hierarchical theory predicts, then its potency should be substantially
reduced if those associations are replaced by different ones: thus an occasion setter
11

should be specific to a CS/US combination. In fact, evidence on this is mixed: Holland
(1989c) trained rats on two feature-negative tasks with two USs (cs1→US1 / OS1:
cs1→nothing / cs2→US2 / OS2: cs2→nothing), and then the rats received pairings of cs1
with US2, and cs2 with US1. Transfer of occasion-setting to these new cs1→US2, and
cs2→US1 associations was complete — response rates to cs1 signalled by OS1, and to cs2
signalled by OS2, were as low after cs1 and cs2 retraining as before — despite these event
combinations never having been subject to occasion setting. This is inconsistent with the
hierarchical account. However, other studies using different techniques have provided
support for specificity to particular CS/US combinations. Bonardi & Ward-Robinson
(2001) trained pigeons on a switching task (Asratyan, 1961) with two USs: OSA signalled
that cs1 was followed by US1, and cs2 by US2, while OSB signalled the opposite (see
Table 5). OSA and OSB were diffuse auditory and visual cues, cs1 and cs2 keylights, and
US1 and US2 were red and white lentils. Then birds received Same trials, on which OSA
and OSB signalled the same CS/US relations and a further keylight, S, was interposed
between cs1 and cs2 and the USs, and also Different trials, which were identical except
that S was replaced by keylight D, and the original CS/US relations were reversed. At
test there was less responding to S than to D, a result which can be understood as a type
of blocking. If OSA and OSB modulated specific CS/US pairings, on Same trials this
would ensure that US delivery was fully predicted by cs1 and cs2, and thus that
acquisition of associative strength by S would be blocked. But on Different trials A and
B would not allow cs1 and cs2 to predict the outcomes that were delivered, making them
surprising and thus able to support learning about D.
Table 5 about here
Bonardi (2007) reported analogous results in a feature-negative task: rats were
trained that cs1 and cs2 were reinforced with US1 and US2 respectively, except when they
12

were signalled by OS (see Table 6). cs1 and cs2 were auditory, US1 and US2 oil and
sucrose, and OS visual. Then the rats were trained on two feature-positive tasks with OS
as the feature. In Group Same OS signalled that cs1 was followed by US1, and cs2 by US2,
but in Group Different these pairings were reversed. If in stage 1 animals learned that
OS signalled specific cs1 → no US1 and cs2 → no US2 associations, then learning should
be more difficult in Group Same, who unlike Group Different had to learn exactly the
opposite of what they had learned in stage 2. That is what was observed (Figure 7).
Table 6 about here
Figure 7 about here
3.3 Perceptual interactions Another factor that discriminates between hierarchical and
configural theory relates to the conditions fostering occasion-setting. Configural
accounts assume that configural cues form most readily when there is a possibility for
perceptual interaction between the to-be-configured cues - yet the conditions promoting
formation of occasion setters are often not those that would be likely to facilitate such
interaction (Holland, 1992). For example, configuring in a feature-positive task seems
more likely when feature and target are presented simultaneously rather than serially. Yet
Holland (1989b) demonstrated that in a feature-positive task with simultaneous OS and
CS presentation (OS & cs→ US / cs→ nothing), the OS behaved more like a simple
Pavlovian CS; its ability to promote responding to the CS transferred well to other CSs
and was attenuated by counterconditioning. But if the OS preceded the CS during
training (OS→cs→US / cs→nothing), it did not transfer to other cues and was not
affected by counterconditioning - it was more like an occasion setter. Occasion setting
was thus fostered by serial training procedures that were less likely to support configural
learning. Configuring also seems more likely when OS and CS are of the same modality
13

as this would allow perception of the two cues to interact more effectively. Yet Holland
(1989b) found that simultaneous positive-patterning tasks (OS & cs→US / cs→nothing /
OS→nothing) were easier to learn when OS and CS were of the same modality, whereas
serial positive-patterning tasks (OS→cs→US / cs→nothing / OS→nothing) were easier
if OS and CS were of different modalities. Thus the serial tasks that foster occasion
setting were easier to learn if configuring was less, rather than more, likely. Conversely,
the simultaneous tasks that were less likely to result in occasion setting were easier to
learn when configuring was more likely.
Other evidence from our laboratory, which addressed the ability of occasion
setters to signal when a US is delivered (Bonardi & Jennings, 2007), casts further doubt
on this perceptual interaction view. Rats were trained on a switching task with two
occasion setters, OS1 and OS2, and two target stimuli cs1 and cs2. OS1 and OS2 were 10s in
duration, and followed by cs1 or cs2, OS and cs being separated by a 5-s trace interval;
cs1 and cs2 were both presented in long and short trials, 30-s and 6-s in duration
respectively, giving four trial types, cs1 -short, cs1 -long, cs2 -short, cs2 -long. All CS
presentations were immediately followed by food, but the OSs gave information about
the delay to food delivery: OS1 signalled cs1 -short and cs2 -long trials, while OS2
signalled cs1 -long and cs2 -short trials (OS1: cs1-short→US / OS1: cs2-long→US / OS2:
cs1-long→US / OS2: cs2-short→US). Responding on probe trials (identical to training
trials except that both cs1 and cs2 were presented for 90s and no US was delivered)
indicated the rats had learned this task: they showed a peak of responding at around 6s
after CS onset on short trials, and 30s after CS onset on long trials. These results could
be compatible with a configural account if the trace of each OS and CS decayed in some
time-dependent manner after its onset. Thus the rats might learn that on long trials on
which OS1: cs2 and OS2: cs1 were reinforced, a configural cue, of the trace of the OS 35s
after its offset plus the trace of the CS 30s after its offset, was paired with food, and
could control timed responding. We reasoned that these temporally sensitive configural
14

cues would be disrupted if we altered the trace interval between the feature and the
target. Thus in a subsequent test we compared responding on training trials with that on
trials in which the interval between OS offset and CS onset was increased from 5 to 29s.
The rats continued to respond at the 'correct' time after CS onset - despite the fact that
the OS trace at this point would have been completely different to that present during
training. This is more consistent with a hierarchical-type account which allows each OS
to signal a specific temporal relationship with a specific CS and outcome delivery.
3.4 Summary Occasion setters are not only specific to the CS, but also to the US with
which they were trained. As to whether the occasion setter acts on a specific
combination of a CS and US, the evidence is mixed; while transfer studies do not
support this prediction, other types of task do. In combination with findings from
operant tasks examining the specificity of discriminative stimuli and inhibitors to
specific response-reinforcer associations (Colwill & Rescorla, 1990; Colwill, 1991;
Bonardi & Hall, 1994b), these data confirm the view that occasion setters are
association-specific in their action. This pattern of findings is not anticipated by
configural theory, but is consistent with the hierarchical account. A further problem for
configural theory stems from its implicit assumption that configuring is more likely
when the to-be-configured cues may interact perceptually. In fact it seems that the more
likely such interaction is, the less likely occasion setting is to occur. Also occasion-
setting-like behaviour can be maintained even when the potential configural cues are
severely degraded. Thus, although the parsimony of a configural account is appealing,
standard versions of such accounts have difficulty dealing with the empirical findings.
4. Mechanisms of occasion-setter formation
15

In comparison with the body of work on transfer, relatively few studies have
explored the learning process by which occasion setting is established. Configural
accounts predict that formation of occasion setters should obey normal associative rules,
provided the to-be-associated stimulus is the configural cue of OS and CS. But because
the hierarchical account assumes that occasion setters act non-associatively, there is no
reason why their formation should be governed by associative rules - and it is not clear
what the alternative should be.7 One starting point was suggested by Mackintosh, who
interpreted Skinner's original formulation by suggesting that the OS (or an operant Sd)
controls operation of CS→US or (R→US) through an associative-type process: "If the
Sd provides subjects with information about the relationship between their actions and
consequences, this is because it is associated not with those actions, nor with their
consequences, but with a representation of the relationship between them." (Mackintosh,
1983; pp.110-111). Thus although the OS's effect on behaviour is, by definition, not
mediated by an association with either CS or US, its ability to exert this control is
nonetheless the product of an association with the 'relationship' between them. Although
this notion has persisted (e.g., Bonardi, 1996; Bonardi et al., 2012), it is not well
specified; nonetheless, the assumptions on which it relies can be tested. For example, it
predicts that the process by which an occasion setter forms should obey standard
associative rules. Thus, assuming that the 'event' to be associated with the OS is the
CS→US pairing, acquisition of occasion setting should show blocking. A corollary of
this is that the CS→US association can be regarded as a unitary 'event' that can enter
into associations. We will consider evidence relating to both of these proposals.
4.1 Learning rules: Bonardi (1991) trained rats that during 3-min presentations of OSA,
5-sec presentations of cs1 were followed by a food US; in OSA's absence cs1 was
7 the exception is Rescorla's US modulation account; in a parallel with conditioned
inhibition, he proposed that occasion setters were established in the presence of the
reinforcement of a CS that possessed inhibitory strength.
16

nonreinforced (OSA: cs1 →US / cs1 →nothing). In stage 2 animals were still trained with
OSA and cs1, but OSA was presented in compound with OSB (OSAOSB: cs1→US /
cs1→nothing); OSB in compound with OSC also signalled reinforcement of cs2 (OSBOSC:
cs2→US / cs2→nothing). Finally the ability of OSB to elevate responding to cs1 and cs2
was evaluated. If in stage 1 OSA became associated with cs1→US, this would block
formation of an association between OSB and cs1→US in stage 2; but as OSC was novel,
cs2→US pairings would not be predicted by OSBOSC, and OSB's association with
cs2→US would be unimpaired. If OSB's ability to act as an occasion setter, promoting
responding to cs1 and cs2, was based on its association with cs1→US and cs2→US, then
OSB should elevate responding to cs2 more than to cs1 - which is what was observed.
Critically this cannot be due to blocking of the Pavlovian association between OSB and
food, as this would affect responding to cs1 and cs2 equally. Configural theory could also
explain these results via blocking. Generalisation between OSA: cs1 and OSAOSB: cs1
could curtail acquisition of associative strength by OSAOSB: cs1, reducing generalisation
to OSB: cs1, which would thus elicit less responding at test than OSB: cs2
Related evidence has been generated in operant tasks. Colwill & Rescorla (1990)
trained rats on two operant contingencies signalled by OSx (OSx: R1→US1 / R1→nothing/
OSx : R2→US2 / R2→nothing). In further training OSx was accompanied by either S or D.
During OSx & S trials the same operant contingencies were in operation as before (OSx &
S: R1→US1 / R1→nothing / OSx & S: R2→US2 / R2→nothing) whereas during OSx & D
trials each response was paired with the alternative outcome (OSx & D: R1→US2 /
R1→nothing / OSx & D: R2→US1 / R2→nothing). At test D was better able than S to elicit
R1 and R2. This is interpretable as blocking by the occasion setter. Initially OSx becomes
associated with R1→US1 and R2→US2. In stage 2, during OSx & S, R1→US1 and R2→US2
are fully predicted by OSx, blocking the ability of S to become associated with them. But
when R1 and R2 are paired with the alternative outcome during OSx & D, the resultant
contingencies are surprising, so D may become associated with them. Again configural
17

theory would treat this as a case of unblocking - in stage 2 the outcomes following OSx
& S and R1, for example, are better predicted than those following OSx & D and R1.
4.2 Summary Occasion-setter formation appears to conform to associative rules, as it
shows blocking: the ability of OS1 to acquire control over cs1→US1 is blocked if a
pretrained occasion setter for cs1→US1 is present. For the reasons we have outlined, this
is not predicted by the hierarchical account unless it assumes that (i) each CS/US pairing
may act as a unitary outcome independent of its constituents, and (ii) OSs acquire their
occasion-setting properties through becoming associated with the CS/US pairing.
5. Elaborations of hierarchical and configural theory
5.1 Elaborated hierarchical account Incorporation of hidden units into the hierarchical
structure can help accommodate these assumptions. Bonardi et al. (2012) suggested an
elaboration of the theory which assumes that when a CS→US association is established,
a hidden unit is recruited that is specific to that association and linked to it via an
asymmetric link. When the association is active, the hidden unit is activated in a normal
associative manner (Figure 8 top panel; upward arrow), and any other stimulus that is
present becomes associated with the hidden unit via standard associative learning;
conversely activation of the hidden unit gates the ability of the CS to predict the US
(Figure 8 top panel; round-headed arrow). Thus a cue that is present when the CS-US
association is established becomes associated with its hidden unit, allowing presentation
of that cue to facilitate flow of activation from CS to US, and act as an occasion setter8.
8 One possibility is that during CS→US formation, when the US is surprising, the
hidden unit is recruited into A1 (cf. Wagner, 1980), and can support learning both as a
'CS' and as a 'US'; once the CS→US association is formed, and the CS successfully
predicts the US, the hidden unit enters A2, and can support learning as a CS, not a US.
18

Figure 8 about here
This modified theory can explain blocking of occasion setting (Bonardi, 1991):
in stage 1 an association linking OSA to the hidden unit modulating cs1 →US forms and
reaches asymptote. In stage 2 this hidden unit is now fully predicted by OSA, so
formation of an association between the added OSB and the hidden unit is blocked, and
OSB cannot acquire any occasion-setting properties for cs1 →US. In contrast, when OSB
and OSC signal the new link cs2→US, a new hidden unit is recruited, with which both
OSB and OSC can become associated. Thus OSB will control responding to cs2 but not cs1.
5.2 Nature of the association representation A second corollary of this modified
hierarchical account relates to whether a particular combination of events can enter into
further associations. Evidence on this point has emerged from studies of acquired
equivalence - the observation that the functional similarity between two cues can be
enhanced if they share a common training history. Honey and Hall (1989) paired
stimulus A with food, and then subsequently with shock. They observed that the fear CR
elicited by A generalised more readily to a second stimulus B if it had also been paired
with food. They attributed this to mediated conditioning (Holland, 1981; Ward-Robinson
& Hall, 1999): when A was paired with shock the representation of the food, with which
it had previously been associated, was activated and became associated with the shock
via a mediated conditioning process. Thus B, also able to activate the food
representation, could also elicit fear at test via this food→shock association. This logic
has been applied to occasion-setting. Honey and Watt (e.g. 1998; cf. Bonardi &
Jennings, 2009) trained rats on a switching task with four, 10-s auditory occasion setters,
OS1, OS2, OS3 and OS4, and two 10-s visual CSs, cs1 and cs2. When preceded by either
OS1 or OS3, cs1 was reinforced with food and cs2 was not, but when preceded by OS2 or
OS4 these contingencies were reversed (see Table 7). Thus OS1 and OS3 signalled the
19

same contingencies between cs1, cs2 and food as did OS2 and OS4. Then OS1 was paired
with shock while OS2 was nonreinforced, and finally fear of OS3 and OS4 was evaluated.
In a parallel of the acquired equivalence result described above, they found greater fear
of OS3, trained in the same manner as the shock-reinforced OS1, than OS4.
Figure 9 about here
These results can also be explained by the elaborated hierarchical account, as it
predicts that the hidden units relating to the cs1→USfood and cs2→nothing pairings will
become activated during the shock conditioning stage. During initial training both OS1
and OS3 become associated with the hidden units linked to the cs1→USfood (Figure 9) and
cs2→nothing links. When OS1 is then paired with shock, the hidden units linked to both
cs1→USfood and cs2→nothing are activated and become associated with shock9. When
OS3 is presented at test, as it can also activate these cs1→USfood and cs2→nothing hidden
units, it also indirectly activates the representation of shock and elicits fear.
This implies that the cs1→USfood and cs2→nothing pairings should themselves
become aversive as a result of this training, as use of these associations will activate
their respective hidden units, and hence the shock representation. This prediction was
tested in two further studies (Bonardi & Jennings, 2009). Rats were given the same
training as in Honey & Watt's (1998) study, but in a final test fear of the cs1→USfood and
cs2→nothing pairings was evaluated: rats were trained to respond for food on an operant
baseline, and while they were responding they experienced presentations of cs1 and cs2;
in one study both were paired with food, and in the other both were nonreinforced. The
rate of operant responding after each of these pairings was compared to baseline rates of
responding before CS onset. If the above analysis is correct, animals should show more
9 In terms of the suggestion made above, if use of the association recruits the hidden unit
into A2, it can enter further learning as a CS, and thus become associated with shock.
20

fear after cs1→USfood than after cs2→USfood trials, but conversely more fear after
cs2→nothing than cs1→nothing trials. This was what was found (Figure 10).
Table 7 about here
Figure 10 about here
5.3 Extended configural theory This elaboration of the hierarchical account has
allowed it to expand its range of predictions, and a similar strategy was applied to
configural theory. Honey and Watt proposed an 'extended' configural theory (e.g. Honey
& Watt, 1999; Honey & Ward-Robinson, 2002) to allow encoding of the US
representation, and hence account for US specificity effects. It assumes that configuring
occurs not through perceptual interaction, but via hidden units (e.g. Pearce, 1994). An
associative structure is supposed comprising input and output units with an intervening
set of hidden units. A given CS or US will uniquely correspond to a specific input or
output unit respectively, but CS/US pairings can also recruit a hidden unit that links their
input and output units (Figure 8 bottom panel). This effectively allows encoding of the
US in the configural cue: for example, if OS1 signals that cs1 predicts US1, OS1 and cs1
will become associated with the same hidden unit p that is also linked to the US1
outcome (Figure 11a). This explains why OS1 might show US-specificity, and transfer
better to cs2 if it is associated with US1 than with US2. Responding will occur to the
extent that the OS1: cs2 compound can activate a hidden unit; thus if cs2 is paired with
US1 it will also be linked to the critical hidden unit p, and elicit more responding than if
it were paired with US2 and thus linked to a second hidden unit q. Note that this
prediction assumes that two sources of activation to p have a greater effect than one
source of activation to each of p and q (Figure 11a (i) and (ii)). If such hidden units are
recruited during normal conditioning with a single CS, then this leads to the prediction
21

that responding should always be greater to a compound of two CSs that predict the
same US than when they predict different USs. Evidence on this is mixed, however:
some studies find more Pavlovian summation when the predicted USs match (Rescorla,
1999), others more responding when they differ (Watt & Honey, 1997), and others no
difference (Ganesan & Pearce, 1988). Nor can this account clearly predict CS/US
specificity: if OS1 signals cs1 →US1, and OS2 signals cs2 →US2, hidden units p and q
form (Figure 11b). If cs1 were then paired with US2, and cs2 with US1 (dotted arrows) the
model would presumably predict that both CSs will become associated with both hidden
units; thus without further assumptions there is no reason to predict that OS1: cs1 would
be any different in its effect on behaviour than OS1: cs2.
Figures 11a and 11b about here
Because it is an associative model, the extended configural account can explain
why occasion-setter formation obeys associative-type principles, provided the rules
governing association of input and hidden units are subject to standard cue competition.
For example, blocking of occasion setting could be explained (Bonardi, 1991): if OSA
and cs1 are linked to the same hidden unit as the US, then as long as the same hidden unit
is recruited in both stages of training, OSA will be able to block acquisition of associative
strength by OSB for this hidden unit, impairing OSB's ability to become an occasion
setter. The model could also explain the results reported by Honey and Watt (1998) and
Bonardi & Jennings (2009; Figure 12). Because both OS1 and OS3 signal that cs1 predicts
USfood, all these events will become associated with p, and so on. When OS1 is
subsequently paired with shock, p is activated and also becomes associated with shock;
thus anything that can also activate p - such as OS3 - will also access the shock
representation and elicit fear. However, explaining Bonardi & Jennings' (2009) results
requires the additional assumption that presentation of USfood can also activate the hidden
22

unit via a feedback mechanism (Figure 12, dotted arrows). Thus presentation of cs1 and
USfood provides two sources of activation to the p hidden unit associated with USshock,
while presentations of cs2 and USfood each activate only one hidden unit that is linked to
OS1 (p and q); thus as two sources of activation to one hidden unit have a greater effect
than one source of activation to two different hidden units, presentations of cs1 and
USfood will produce more fear than presentations of cs2 and USfood.
Figure 12 about here
5.4 Summary Extended configural theory overcomes the inability of more traditional
configural theories to explain US specificity, but at some expense. Current formulations
do not specify the conditions required for recruitment of hidden units, and the account
must also make additional assumptions - for example, about the effect of summing two
sources of activation to one hidden unit, or how US presentation can also activate its
associated hidden unit. Elaborated hierarchical theory must also make assumptions,
about the properties of the associations linking the association to its hidden unit, the
rules governing how the hidden unit is activated and when it can undergo learning.
Given this potential impasse, a different means is required to discriminate between these
approaches. We now turn to some different classes of evidence that might achieve this.
6. Further discriminating evidence
6.1 OS/CS symmetry One fundamental distinction between the elaborated hierarchical
and extended configural theories lies in the role of the occasion setter. Configural
theories assume only the processes of associative learning: activity in a stimulus
representation can only influence activation of a second stimulus via an association
between them. Thus there is no qualitative distinction between the action of the OS and
23

CS, because there is only one way that one stimulus can interact with another10. In
contrast hierarchical theory proposes that CS representations can also interact in a
nonassociative manner, that allows an occasion setter to facilitate the flow of activation
in an associative link. Thus the hierarchical account assumes that the OS plays a
qualitatively distinct role from that of a CS in controlling behaviour.
We examined this distinction in two studies (Bonardi et al., 2012). Rats were
trained on two positive-patterning discriminations with two occasion setters OS1 and
OS2, two conditioned stimuli cs1 and cs2 and two USs US1 and US2 (Table 8). According
to extended configural theory this should result in formation of two hidden units, one
linking OS1 and cs1 to US1, and another linking OS2 and cs2 to US2 (cf. black arrows
Figure 11b). The OSs were visual, the CSs auditory, and the USs sucrose and oil.
Stimulus presentations were of 10-s duration, and OS offset was separated from CS
onset by a 5-s trace interval. The rats were also trained with two types of transfer
stimulus (i) two separately trained test excitors cs3 and cs4 which, like the CSs, were
auditory and were immediately followed by either US1 or US2; and (ii) two separately
trained pseudo-occasion setters POS1 and POS2 which, like the OSs, were visual and
were paired with either US1 or US2 after a 15-sec trace interval.
Table 8 about here
Figure 13 about here
The studies differed in their test procedure. In one we examined performance on
trials that were identical to the compound training trials except that cs1 and cs2 were
replaced by one of the test excitors, cs3 or cs4. The constituents of these test compounds
10 Although earlier theories, as we saw above, allowed for inhibitors to modulate
activation of representations rather than activating anything directly, later authors have
rejected this possibility (Mackintosh, 1983).
24

either both signalled the same outcome (OS1: cs3 / OS2: cs4) or different outcomes (OS1:
cs4 / OS2: cs3). Consistent with the US specificity predicted by both elaborated
hierarchical and extended configural theory, we found more responding to cs3 and cs4 on
same than on different trials. In the second study the test trials were identical to the
compound training trials except this time OS1 and OS2 were replaced by one of the
pseudo-occasion setters POS1 and POS2. These trials were again classified as either
same (POS1: cs1 / POS2: cs2) or different, (POS1: cs2 / POS2: cs1). According to extended
configural theory, as there is no qualitative distinction in the ability of OS and CS to
activate the hidden unit, replacing either element of the trained compound will produce
the same effect, more responding on same than different trials; but this was not observed.
Mean rates of responding on same (s) and different (d) trials in each of the 2-trial blocks
of each test were converted to a ratio of form s/d and averaged; ratios greater than 1
indicate more responding on same than different trials. The data, in Figure 13, suggest
that the ratios exceed 1 in the first study (p = 0.027), but not in the second; ANOVA
revealed a significant effect of experiment, F(1, 30) = 4.74, MSe = .493, p = .04.
We interpreted these findings as supporting the hierarchical view that OS and CS
do not play functionally equivalent roles. Although the CS acts in a standard associative
manner, the OS does not. Thus replacing the CS of the training compound with a
separately trained stimulus retains the occasion-setting nature of the task in a way that
replacing the OS with a separately trained pseudo-occasion setter does not, meaning the
hierarchical account need not predict the same pattern of responding in both tests. In
contrast, the extended configural theory must predict that both tests are functionally the
same: both OS1 and CS1 are linked to the same hidden unit, and so replacing either of
these cues should produce more responding if the replacement cue is associated with the
same US as the cue that remains. (Moreover, we noted above that this prediction relies
on specific assumptions about the activation thresholds for the hidden units, which
means it has difficulty explaining the finding that summation is greater when the pair of
25

CSs predict two different USs than when they predict the same US (Watt & Honey,
1997). It is thus of interest that in the final study of the series we combined simple CSs
predicting the same or different outcomes, and found less responding when the USs
matched11.) A dissociation in the pattern of responding to CS compounds whose
components predict the same, or different, outcomes depending on whether one of those
components is an occasion setter does not support the extended configural account.
6.2 Parsing of the association Another study attempted to discriminate the two theories
in a different way. In a task similar to that described above two groups of 8 rats received
training trials in which cs1 and cs2, 10-sec presentations of a 74-dB,10-Hz clicker, or the
illumination of the tray light, were each paired with one of two USs; for half the rats the
click was paired with sucrose and the traylight with oil, and for the rest the reverse. They
also received nonreinforced compound trials on which a third cue OS, the 10-s
illumination of a 2.8-W jewel light, signalled simultaneous presentation of cs1 and cs2
after a 5-s trace interval (OS: cs1&cs2 →nothing / cs1 →US1 / cs2 →US2).12 The data
from the discrimination training phase are presented in Figure 14; ANOVA showed that
both groups learned the discrimination, F(1, 14) = 17.82, p = .001; smallest p involving
the group factor: F(1, 14) = 1.06, p = .32. Then, to equate pairings of cs1 and cs2 with the
two USs before the test, rats received reverse training in which cs1 was paired with US2
and cs2 with US113. The groups did not differ in this phase, and responding remained high
(20.1 and 24.9 rpm for Groups Same and Different respectively on the final 2-session
block; see below); ANOVA with group and 2-session block as factors revealed nothing
11 We tested the effect of signalling cs1 and cs2 with cs3 and cs4, again either in same
(cs3→cs1 and cs4→cs2) or different (cs3→cs2 and cs4→cs1) combinations, and found
significantly higher responding on different trials in the first test block, p = 0.024.
12 There were 56 trials scheduled per session, with a variable inter-trial interval (mean
of 75s, range of 60-90s), and 26 training sessions in this stage; each comprised 4 cs1
→US1 and 4 cs2 →US2 trials, and in all but the first 10 sessions also 48 OS: cs1&cs2
→nothing trials.
13 This training comprised 18 further sessions, each of 8 cs1 →US2 and 8 cs2 →US1
trials.
26

significant; the smallest p was associated with the interaction, F(8, 112) = 1.07, p = .07,
and the effect of group was not significant, F < 1.
Figure 14 about here
Table 9 about here
Finally the rats were divided into two groups, and trained on two feature-positive
discriminations. In Group Same the OS signalled reinforcement of cs1 with US1, and cs2
with US2, the opposite of what had been the case during the initial feature-negative
training when OS signalled cs1 would not be followed by US1, and cs2 would not be
followed by US2. In contrast, for Group Different the CS→US pairings during feature-
positive training were reversed, so test training did not directly contradict the feature-
negative discrimination in this way14.
The extended hierarchical account assumes that during feature-negative training,
cs1→US1 and cs2→US2 associations form, and that one hidden unit is recruited for each.
Although it does not specify how negative occasion setters form, let us assume for
simplicity that a stimulus that is present when cs1 is paired with the unexpected omission
of US1 acquires modulatory control over a cs1→ noUS1 association, which reduces the
extent to which cs1 can predict US1. Thus the OS would acquire inhibitory control over
the cs1→US1 and cs2 →US2 associations, and this control would not be compromised
when different cs1→US2 and cs2 →US1 associations were subsequently formed. Thus
Group Same, who had to learn the opposite of the relations in operation during feature-
14 There were two phases in this stage; the first, in which the feature-positive
discriminations were acquired, comprised 10 sessions, each of 24 OS →nothing, 12 cs1
→nothing, 12 cs2 →nothing, and for Group Same 4 OS: cs1 →US1 4 OS: cs2 →US2
trials; the second, in which performance was tested, comprised 12 sessions identical to
those of the first phase except that there could be 2, 4 or 8 OS: cs1 →US1 and OS: cs2
→US2 trials per session. For Group Different US1 and US2 were reversed.
27

negative training at test, would perform less effectively at test than Group Different, as
only the former group would have to overcome the inhibitory modulation over these
specific associations that was established in the feature-negative training stage.
Extended configural theory instead predicts that initial negative-patterning
training would result in hidden unit p linking cs1 and US1, and q linking cs2 and US2. The
model assumes that learning about non-reinforced trials is governed by an excitatory
association from the hidden unit to a no-US representation (e.g., Allman, Ward-
Robinson & Honey, 2004; cf. Konorski, 1967). Thus OS, cs1 and cs2 should be equally
associated with a no-US1 representation via hidden unit r, and a no-US2 representation
via the hidden unit s (Figure 15). When cs1 and cs2 are each paired with the alternative
USs, they will become linked to q and p respectively, resulting in the associative
structure shown in Figure 15. It is clear from the symmetry here that there would be no
good reason to predict why animals in Group Same should learn their feature-positive
test discrimination any less effectively than those in Group Different.
Figure 15 about here
The results of this experiment may be seen in Figure 16. Levels of responding
declined gently over the course of testing, presumably because the number of reinforced
trials per session was on average lower than it had been in the previous phases (see
footnote 12). While Group Different performed accurately, responding more on reinforced
trials on which cs1 and cs2 were signalled by OS than when cs1 and cs2 were presented
alone, Group Same showed the opposite pattern. ANOVA with group, trial type
(reinforced and nonreinforced) and blocks as factors revealed a significant interaction
between group and trial type, F(1, 14) = 7.47, p = .02, and simple main effects revealed
that the effect of trial type was significant in Group Different, F(1, 14) = 6.37, p = .02,
28

but not in Group Same, F(1, 14) = 1.80, p = .20. The results were, therefore, in accord
with the predictions of the hierarchical account.15
Figure 16 about here
7. Conclusions: do we need a hierarchical theory?
We have reviewed theories of occasion setting which fall into two classes, those
that require modification of existing associative learning principles, and those that do
not. In the former category the US modulation and memory systems accounts were
largely unsupported by evidence on the CS- and US-specificity of occasion setting,
leaving the hierarchical account as the main contender. Configural theories fall into the
second class, but are unable to explain why occasion setters are US-specific without
considerable modification, resulting in extended configural theory. Conversely,
hierarchical theory could not explain the apparent constraints on occasion-setter
formation without elaboration of its associative structure. Both theories must therefore
make a number of assumptions to account for the existing data, and more precise
specification of the conditions required for recruitment of hidden units is required,
especially for extended configural theory. Future work could usefully compare the
degree to which these two relatively informal approaches can be expressed as
computational models, and explore any concrete predictions that are generated in this
way. Equally, the extent to which the more formal models that have been developed (cf.
15 It should be noted that there is an inevitable confound in this experiment, as the
cs1→US1 and cs2 →US2 associations were always learned before the cs1→US2 and cs2
→US1 associations. Thus the occasion-set relations of the test were learned first for
Group Same but second for Group Different, and this could also be responsible for the
differences at test. However, if the reverse-order training preceded the negative
occasion-setting training there would be no guarantee that OS acquired negative
occasion-setting properties over cs1→US1 and cs2 →US2 and not to some extent over
cs1→US2 and cs2 →US1 - fatally compromising the logic of the experiment.
29

Kutlu & Schmajuk, 2012; Vogel, Ponce & Wagner, this volume) can accommodate the
key classes of evidence described here could be explored. An alternative means of
discriminating the two approaches could focus on their underlying assumptions - the
theories differ fundamentally in how they conceptualise the roles of OSs and CSs - and
this may provide grounds for discriminating between them. We reported two of our own
studies that address this distinction, and the results were consistent with hierarchical
theory's predictions. It is also possible that neuroscientific approaches may be able to
dissociate the two mechanisms.
Of course these two approaches should not be taken as mutually exclusive. There
is evidence outside the occasion-setting literature that supports the existence of
configural cues - although whether such evidence can be accommodated within the
perceptual interaction version of configural theory, or requires the extended configural
interpretation, is unclear. The question is thus more one of whether an additional notion
of hierarchical control is also required. We have argued in this article that it is, and that
any general associative theory that is able to explain occasion-setting effects will need to
incorporate such a mechanism. In our view the development of properly formalised
mechanisms for hierarchical control, and their incorporation into a more general model,
probably represent the path to a truly comprehensive account of associative and
occasion-setting phenomena.
30

Acknowledgements
This work was funded by the BBSRC (grant number BBS/B/01251).
31

References
Allman, M. J., Ward-Robinson, J., & Honey, R. C. (2004). Associative Change in the
Representations Acquired During Conditional Discriminations: Further Analysis
of the Nature of Conditional Learning. Journal of Experimental Psychology:
Animal Behavior Processes, 30, 118-128.
Asratyan, E.A. (1961). The initiation and localisation of cortical inhibition in the
conditioned reflex arc. Annals of the New York Academy of Sciences, 92, 1141-
1159.
Bonardi, C. (1991). Blocking of occasion setting in feature-positive discriminations.
Quarterly Journal of Experimental Psychology, 43B, 431-448.
Bonardi, C. (1996). Transfer of occasion setting: The role of generalization decrement.
Animal Learning and Behavior, 24, 277-289.
Bonardi, C. (2007). Occasion setting is specific to the CS-US association. Learning and
Motivation, 38, 208-228.
Bonardi, C., Bartle, C., & Jennings, D. (2012). US specificity of occasion setting:
Hierarchical or configural learning? Behavioural Processes, 90, 311-322.
32

Bonardi, C., & Hall, G. (1994a). Occasion-setting training renders stimuli more similar:
Acquired equivalence between the targets of feature-positive discriminations.
Quarterly Journal of Experimental Psychology, 47B, 63-81.
Bonardi, C., & Hall, G. (1994b). Discriminative inhibition is specific to the response-
reinforcer association but not to the discriminative stimulus. Journal of
Experimental Psychology: Animal Behavior Processes, 20, 278-291.
Bonardi, C., & Jennings, D. (2007). Occasion setting of timing behaviour. Journal of
Experimental Psychology: Animal Behavior Processes, 33, 339-348.
Bonardi, C., & Jennings, D. (2009). Learning about associations: Evidence for a
hierarchical account of occasion setting. Journal of Experimental Psychology:
Animal Behavior Processes, 35, 440-445.
Bonardi, C., & Ward-Robinson, J. (2001). Occasion setters: Specificity to the US and
the CS-US association. Learning and Motivation, 32, 349-366.
Bouton, M.E., & Nelson, J.B. (1998). Mechanisms of feature-positive and feature-
negative discrimination learning in an appetitive conditioning paradigm. In N.A.
Schmajuk & P.C. Holland (Eds.) Occasion setting: Associative learning and
cognition in animals. (pp.69-112). Washington D.C.: American Psychological
Association.
33

Brandon, S.E., Vogel, A.H., & Wagner, A.R. (2000). A componential view of configural
cues in generalization and discrimination in Pavlovian conditioning. Behavioral
Processes, 110, 67-72.
de Brugada, I., Garcia-Hoz, V., Bonardi, C., & Hall, G. (1995). The role of stimulus
ambiguity in conditional learning. Journal of Experimental Psychology: Animal
Behavior Processes, 21, 275-284.
Colwill, R.M., & Rescorla, R.A. (1990). Evidence for the hierarchical structure of
instrumental learning. Animal Learning and Behavior, 18, 71-82.
Colwill, R. M. (1991). Negative discriminative stimuli provide in- formation about the
identity of omitted response-contingent outcomes. Animal Learning & Behavior,
19, 326-336.
Cousineau, D. (2005). Confidence intervals in within-subject designs: A simpler
solution to Loftus and Masson’s method. Tutorial in Quantitative Methods for
Psychology, 1, 42-45.
Davidson, T. L., & Rescorla, R. A. (1986). Transfer of facilitation in the rat. Animal
Learning & Behavior, 14, 380-386.
Ganesan, R., Pearce, J.M. (1988). Interactions between conditioned stimuli for food and
water in the rat. Quarterly Journal of Experimental Psychology, 40B, 229–241.
34

Haselgrove, M., Robinson, J., Nelson, A., & Pearce, J.M. (2008). Analysis of an
ambiguous-feature discrimination. Quarterly Journal of Experimental
Psychology, 61, 1710–1725.
Hall, G., & Honey, R. C. (1989). Contextual effects in conditioning, latent inhibition and
habituation: Associative and retrieval functions of contextual cues. Journal of
Experimental Psychology: Animal Behavior Processes, 15, 232-241.
Holland, P.C. (1981). Acquisition of representation-mediated conditioned food
aversions. Learning & Motivation, 12, 1-18.
Holland, P.C. (1983). Occasion-setting in Pavlovian feature positive discriminations. In
M.L. Commons, R.J. Herrnstein, & A.R. Wagner (Eds.), Quantitative analyses of
behavior: Discrimination processes (Vol. 4, pp. 183-206). New York: Ballinger.
Holland, P.C. (1985). The nature of conditioned inhibition in serial and simultaneous
feature negative discriminations. Information processing in animals:
Conditioned inhibition (pp. 267-297). Hillsdale, NJ: Erlbaum.
Holland, P.C. (1986). Temporal determinants of occasion setting in feature-positive
discriminations. Animal Learning and Behavior, 14, 111-120.
Holland, P.C. (1989a). Feature extinction enhances transfer of occasion setting. Animal
Learning and Behavior, 17, 269-279).
Holland, P.C. (1989b). Acquisition and transfer of conditional discrimination
35

performance. Journal of Experimental Psychology: Animal Behavior Processes,
15,154-165.
Holland, P.C. (l989c). Transfer of negative occasion setting and conditioned inhibition
across conditioned and unconditioned stimuli. Journal of Experimental
Psychology: Animal Behavior Processes, 15, 311-328.
Holland, P.C. (1992). Occasion setting in Pavlovian conditioning. In D. Medin (Ed.) The
Psychology of Learning and Motivation. (Vol. 28 pp.69-125). San Diego:
Academic Press.
Honey, R.C., & Hall, G. (1989). Acquired equivalence and distinctiveness of cues.
Journal of Experimental Psychology: Animal Behavior Processes, 15, 338-346.
Honey, R.C., Hall, G., & Bonardi, C. (1993). Negative priming in associative learning:
Evidence from serial conditioning procedures. Journal of Experimental
Psychology: Animal Behavior Processes, 19, 90-97.
Honey, R.C., & Ward-Robinson, J. (2002). Acquired equivalence and distinctiveness of
cues: I. Exploring a neural network approach . Journal of Experimental
Psychology: Animal Behavior Processes, 28, 378-387.
Honey, R.C., Watt, A. (1998). Acquired relational equivalence: Implications for the
36

nature of associative processes. Iournal of Experimental Psychology: Animal
Behavior Processes, 24, 325–334.
Honey, R.C., Watt, A. (1999). Acquired relational equivalence between contexts and
features. Journal of Experimental Psychology: Animal Behavior Processes, 25,
324–333.
Jenkins, H.M. (1985). Conditioning inhibition of keypecking in the pigeon. In R. R.
Miller & N. E. Spear (Eds.), Information processing in animals: Conditioned
inhibition (pp. 327-353). Hillsdale, NJ: Erlbaum.
Konorski, J. (1948). Conditioned reflexes and neuron organisation. Cambridge
University Press.
Konorski, J. (1967). Integrative activity of the brain. Chicago University Press.
Kutlu, M.G., & Schmajuk, N.A. (2012). Solving Pavlov's puzzle: Attentional,
associative, and flexible configural mechanisms in classical conditioning.
Learning & Behavior, 40, 269-291.
Mackintosh, N.J. (1983). Conditioning and Associative Learning. Oxford University
Press.
37

Morell, J.R., & Davidson, T.L. (2002). Transfer across unconditioned stimuli in serial feature
discrimination training. Journal of Experimental Psychology: Animal Behavior
Processes, 28, 83-96.
Pearce, J.M. (1987). A model for stimulus generalization in Pavlovian conditioning.
Psychological Review, 94, 61-73.
Pearce, J.M. (1994). Similarity and discrimination: A selective review and a
connectionist model. Psychological Review, 101, 587-607.
Pearce, J.M., George, D.N., & Aydin., A. (2002). Summation: Further assessment of a
configural theory. Quarterly Journal of Experimental Psychology, 55B, 61–73.
Rescorla, R.A. (1972). Evidence for "Unique stimulus" account of configural
conditioning. Journal of Comparative and Physiological Psychology, 85, 331-
338.
Rescorla, R. A. (1969). Pavlovian conditioned inhibition. Psychological Bulletin, 72, 77-
94.
Rescorla, R.A. (1985). Conditioned inhibition and facilitation. In R.R. Miller & N.E.
Spear (Eds.), Information processing in animals: Conditioned Inhibition.(pp.
299-326). Hillsdale, NJ: Erlbaum.
38

Rescorla, R.A. (1986). Facilitation and excitation. Journal of Experimental Psychology:
Animal Behavior Processes, 12, 325-332.
Rescorla, R. A. (199Ia). Combinations of modulators trained with the same and different
target stimuli. Animal Learning & Behavior, 19, 355-360.
Rescorla, R. A. (1991b). Transfer of inhibition and facilitation mediated by the original
target stimulus. Animal Learning & Behavior, 19, 65-70.
Rescorla, R.A., (1999). Summation and overexpectation with qualitatively different
outcomes. Learning and Behavior, 27, 50–62.
Ross, R.T., & Holland, P.C. (1981). Conditioning of simultaneous and serial feature-
positive discriminations. Animal Learning and Behavior, 9, 293-303.
Skinner, B.F. (1938). The behavior of organisms. New York: Appleton-Century-Crofts.
Swartzentruber, D. (1995). Modulatory mechanisms in Pavlovian conditioning. Animal
Learning and Behavior, 23, 123-143.
Vogel Ponce Wagner Ref
Wagner, A. R. (1981). SOP: A model of automatic memory processing in animal
behavior. In N. E. Spear & R. R. Miller (Eds.), Information processing in
39

animals: Memory mechanisms (pp. 5-47). Hillsdale, NJ: Erlbaum.
Ward-Robinson, J., & Hall, G. (1999). The role of mediated conditioning in acquired
equivalence. Quarterly Journal of Experimental Psychology B, 52B, 335-350.
Watt, A., Honey, R.C. (1997). Combining CSs associated with the same or different
USs. Quarterly Journal of Experimental Psychology, 50B, 350–367.
Williams, D.A., Sagness, K.E., & McPhee, J.E. (1994). Configural and Elemental
Strategies in Predictive Learning, Journal of Experimental Psychology:
Learning, Memory, and Cognition, 20, 694-709.
40

TRAIN TEST
Same Different Target
OS1: cs1 → USfood
cs1 → nothing
OS2: cs2 → USfood
cs2 → nothing
OS1: cs3 → USfood
cs3 → USfood
OS2: cs4 → USfood
cs4 → USfood
OS1: cs1 OS1: cs2 cs1
OS2: cs2 OS2: cs1 cs2
OS1: cs3 OS1: cs4 cs3
OS2: cs4 OS2: cs3 cs4
Table 1: Design of Bonardi (1996). OS1 and OS2 were 10-s presentations of diffuse
auditory or visual stimuli; cs1, cs2, cs3 and cs4 were 5-s presentations of different
keylights, which immediately followed OS presentations on compound trials. USfood
(access to grain) occurred on CS termination on reinforced trials.
41

TRAIN 1 TRAIN 2 REVALUE TEST
OS1: cs1→USfood
cs1→nothing
OS2: cs2→USfood
cs2→nothing
cs3→USfood
OS1: cs1→USfood
cs1→nothing
OS2: cs2→USfood
cs2→nothing
cs3→nothing
cs1→USfood
cs2?
cs3?
Table 2: Design of Bonardi & Hall, 1994a. OS1 was a 3-minute presentation of a visual
cue, cs1, cs2 and cs3 were 10-s auditory cues. Food only followed cs1 and cs2 when they
occurred during OS1 and OS1 respectively; unaccompanied cs1 and cs2 were
nonreinforced. Presentations of cs3 were always unaccompanied.
42

OS TRAIN TRANSFER
TRAIN
TEST
Same Different Target
OS1: cs1→US1
cs1→nothing
OS1→nothing
OS2: cs2→US2
cs2→nothing
OS2 →nothing
cs3→US1
cs4→US2
OS1: cs3 OS1: cs4 cs3
OS2: cs4 OS2: cs3 cs4
Table 3: Design of Bonardi Bartle & Jennings (2012) Experiment 1. OS1 and OS2 were
10-s visual stimuli, cs1, cs2, cs3 and cs4 10-s auditory stimuli, and US1 and US2 sucrose
or oil. There was a 5-s trace interval between OS and CS presentations on compound
trials.
43

TRAIN 1 TRAIN 2 TRANSFER
TRAIN
TEST
Same Different
OS1: cs1→nothing
OS1→nothing
cs1→US1
OS2: cs2→nothing
OS2→nothing
cs2 →US2
OS1: cs1→nothing
OS1→US1
cs1→US1
OS2: cs2→nothing
OS2→US2
cs2→US2
cs3→US1
cs4→US2
OS1: cs3 OS1: cs4
OS2: cs4 OS2: cs3
Table 4: OS1 and OS2 were 10-s visual stimuli, cs1, cs2, cs3 and cs4 10-s auditory stimuli,
and US1 and US2 sucrose or groundnut oil. There was a 5-s trace interval between OS
and CS presentations on compound trials.
44

TRAIN 1 Same Different TEST
OSA: cs1 →US1
OSA: cs2 →US2
cs1 →nothing
cs2 →nothing
OSB: cs1 →US2
OSB: cs2 →US1
OSA: cs1 S →US1 OSA: cs1 D →US2
OSA: cs2 S →US2 OSA: cs2 D →US1
OSB: cs1 S →US2 OSB: cs1 D →US1
OSB: cs2 S →US1 OSB: cs2 D →US2
S
D
Table 5: Design of Bonardi & Ward-Robinson (2001). OSA and OSA were 10-s
presentations of diffuse auditory or visual stimuli; cs1, cs2, S and D were 5-s
presentations of different keylights, and immediately followed OS presentations on
compound trials. US1 and US2 were red and white lentils.
45

TRAIN TEST
cs1 →US1
OS: cs1 →nothing
cs2 →US2
OS: cs2 →nothing
Group Same
OS: cs1 →US1 cs1 →nothing
OS: cs2 →US2 cs2 →nothing
Group Different
OS: cs1 →US2 cs1 →nothing
OS: cs2 →US1 cs2 →nothing
Table 6: Design of Bonardi, 2007. OS was a 10-s visual stimuli, cs1 and cs2 10-s
auditory stimuli, and US1 and US2 sucrose or groundnut oil. There was a 5-s trace
interval between OS and CS presentations on compound trials.
46

TRAIN REVALUE OS TEST ASSOCIATION TEST
OS1: cs1 →USfood
OS1: cs2 →nothing
OS3: cs1 →USfood
OS3: cs2 →nothing
OS2: cs2 →USfood
OS2: cs1 →nothing
OS4: cs2 →USfood
OS4: cs1 →nothing
OS1 →USshock
OS2→nothing
OS3 ?
OS4 ?
cs1→USfood
cs2 →USfood
or
cs1 →nothing
cs2 →nothing
Table 7: Design of Bonardi & Jennings (2009). OS1, OS2, OS3 and OS4 were 10-s
auditory cues, cs1 and cs2 10-s visual cues; cs onset coincided with OS offset on training
trials.
47

TRAIN 1 TRAIN 2 TRANSFER
TRAIN
TEST
Same Different
OS1: cs1 →US1
OS1 →nothing
cs1 →nothing
OS2: cs2 →US2
OS2 →nothing
cs2 →nothing
OS1: cs1 →US1
OS1 →US1
cs1 →nothing
OS2: cs2 →US2
OS2 →US2
cs2 →nothing
cs3→US1
cs4→US2
POS1 →US1
POS2 →US2
OS1: cs3 OS1: cs4
OS2: cs4 OS2: cs3
POS1: cs1 POS1: cs2
POS2: cs2 POS2: cs1
Table 8: Bonardi, Bartle & Jennings (2012) Experiments 3a and 3b: OS1, OS2, POS1 and
POS2 were 10-s visual stimuli, cs1, cs2, cs3 and cs4 10-s auditory stimuli, and US1 and
US2 sucrose or groundnut oil. There was a 5-s trace interval between OS and CS
presentations on compound trials.
48

TRAIN 1 TRAIN 2 TEST
cs1 →US1
cs2 →US2
OS: cs1&cs2 →nothing
cs1 →US2
cs2 →US1
Group Same
OS: cs1 →US1 OS: cs2 →US2
cs1 →nothing cs2 →nothing
Group Different
OS: cs1 →US2 OS: cs2 →US1
cs1 →nothing cs2 →nothing
Table 9: OS was a 10-s visual stimulus, cs1 and cs2 10-s auditory or visual stimuli, and
US1 and US2 sucrose or groundnut oil. There was a 5-s trace interval between OS and
CS presentations on compound trials.
49

!
cs
1
US
1
OS
cs
2
i
ii
US
2
cs
3
Figure 1: According to hierarchical accounts an OS signalling that cs1 will be followed
by US1 acts either on the US1 representation (i; e.g., Rescorla, 1985) or on the cs1→US1
link (ii, e.g., Holland, 1983). The accounts differ in their predictions about the OS’s
action on a stimulus cs2 that has signalled US1; neither predicts that the OS will have an
effect on a stimulus cs3 that signals US2. The pointed arrow indicates an associative
link; round-headed arrow indicates facilitation of (i) activation of US1 representation or
(ii) transmission of activation via cs1→US1 association. For further details, see text.
50

Figure 2: Results of Bonardi 1996 Experiment 2: Group mean response rates to the
occasion-set cs1 and cs2 and the non-occasion-set cs3 and cs4 on same (OS1:cs1, OS2:cs2,
OS1:cs3, OS2:cs4) and different (OS1:cs2, OS2:cs1, OS1:cs4, OS2:cs3) trials or when
presented alone. Error bars show within-subject confidence intervals (Cousineau,
2005). For further details, see text and Table 2.
51

Figure 3: Results of Bonardi & Hall, 1994a Experiment 1. Mean response rates to cs2
(which had been occasion-set) and cs3 (not occasion-set, but conditioned and
extinguished) in the generalisation test. Error bars show within-subject confidence
intervals (Cousineau, 2005). For further details, see text and Table 2.
52

Figure 4: Group mean response rates to cs3 and cs4 on same, different and target trials
in the two test session blocks of Bonardi, Bartle & Jennings (2012) Experiment 1. Error
bars show within-subject confidence intervals (Cousineau, 2005). For further details,
see text and Table 3.
53

Figure 5: Mean response rates during cs1 and cs2 when nonreinforced and signalled by
OS1 and OS2 (OS1: cs1-, OS2: cs2-) or when presented alone and reinforced (cs1 →US1,
cs2→US2), and also during the occasion setters (OS1 and OS2 ) and the trace interval
separating each OS and each cs on compound trials, during the training stage. Error
bars show within-subject confidence intervals (Cousineau, 2005). For further details,
see text and Table 4.
54

Figure 6: Mean response rates to cs3 and cs4 on same, different and target alone trials
during the two test session blocks. Error bars show within-subject confidence intervals
(Cousineau, 2005). For further details, see text and Table 4.
55

Figure 7: Group mean rates of responding to cs1 and cs2 when signalled by OS (OS:cs)
and presented alone (cs) in the test session of Bonardi, 2007 Experiment 2. Error bars
show within-subject confidence intervals calculated separately for each group
(Cousineau, 2005). For further details, see text and Table 6.
56

!
US
OS
cs
US
OS
cs
Figure 8: Top panel: Associative structure according to the elaborated hierarchical
account. Pointed arrows indicate associative links; the round-headed arrow indicates
the facilitation of activity transmission through the associative link; the circle denotes a
hidden unit. Formation of the association recruits the hidden unit.
Bottom panel: Associative structure according to the extended configural account.
Pointed arrows indicate associative links; the circle denotes a hidden unit. For further
details, see text.
57

!
US
food
!
OS
1
!
cs
1
!
OS
3
!
US
shock
!
Figure 9: Associative structure arising from Honey & Watt's (1998) experiment (Table
7), according to the elaborated hierarchical account. Both OS1 and OS3 become
associated with the hidden unit of the cs1→USfood association during initial training.
When OS1 is paired with shock the hidden unit becomes associated with the shock, thus
allowing future presentations of both OS3 and cs1→USfood to activate the shock
representation via the hidden unit. Pointed arrows indicate associative links; the round-
headed arrow indicates the facilitation of activity transmission through the associative
link; the circle denotes hidden unit.
58

Figure 10: Mean suppression ratios for responding after cs1→USfood and cs2→USfood and
cs1→nothing and cs2→nothing in the test sessions of Bonardi & Jennings (2009;
Experiments 1 and 2 respectively). Error bars show within-subject confidence intervals
(Cousineau, 2005) calculated separately for each group. For further details, see text
and Table 7.
59

!
US
1
OS
1
cs
1
p
q
(i)
US
2
cs
2
US
1
p
q
(ii)
US
2
OS
1
cs
1
cs
2
Figure 11a: The extended configural theory account of US specificity. Associative
structures arising in feature-positive discriminations in which OS1 signals reinforcement
of cs1 with US1, and (i) cs2 also predicts US1 or (ii) cs2 predicts US2. Pointed arrows
indicate associative links; circle denotes hidden unit.
!
US
1
p
q
US
2
OS
1
OS
2
cs
1
cs
2
cs
2
cs
1
Figure 11b: Associative structures arising in feature-positive discriminations in which
OS1 signals reinforcement of cs1 with US1, OS2 signals reinforcement of cs2 with US2, and
then cs1 is paired with US2 and cs2 is paired with US1. Pointed arrows indicate
associative links; circles denotes hidden units.
60

!
US
no food
q
US
no food
s
US
food
OS
1
p
OS
3
US
food
r
cs
1
US
shock
US
shock
OS
2
OS
4
cs
2
OS
1
OS
3
cs
2
OS
2
OS
4
cs
1
Figure 12: Associative structure arising from Honey & Watt's (1998) experiment (Table
7). According to the extended configural account, both OS1 and OS3 signal that cs1
predicts USfood , so OS1 OS3 and cs1 become associated with p, which is also linked to
USfood. Pointed arrows indicate associative links; circles denote hidden units. US
presentation can also activate the hidden unit with which it is linked via a feedback
mechanism (dotted arrow).
61

Figure 13: Ratio of same/different responding in Experiments 3a and 3b of Bonardi,
Bartle & Jennings, 2012. In Experiment 3a test trials compared responding during
transfer excitors cs3 and cs4 signalled by the trained occasion setters OS (same OS1: cs3
and OS2: cs4 or different OS1: cs4 and OS2: cs3), and in Experiment 3b during the CSs
from the occasion-setting discrimination cs1 and cs2 signalled by pseudo-occasion setters
POS (same POS1: cs1 and POS2: cs2 or different POS1: cs2 and POS2: cs1). Error bars
show standard error of the mean. For further details, see text and Table 8.
62

Figure 14: Group mean response rates during cs1 and cs2 on reinforced cs1→US1
cs2→US2 and nonreinforced OS: cs1&cs2 →nothing trials. Error bars show within-
subject confidence intervals (Cousineau, 2005) calculated separately for each group.
For further details, see text and Table 9.
63

!
cs
2
p
cs
1
no US
1
r
US
1
q
US
2
no US
2
OS
s
cs
1
cs
2
cs
2
cs
1
cs
2
cs
1
OS
Figure 15: Initial training results in cs1 and cs2 becoming linked to US1 and US2
respectively by two hidden units, p and q. On nonreinforced trials two further hidden
units are established, one linking OS, cs1 and cs2 with noUS1, and another linking these
same stimuli with no US2. Finally cs1 and cs2 are each paired with the alternative
outcomes, becoming linked to q and p respectively. Pointed arrows indicate associative
links; circle denotes hidden unit.
64

Figure 16: Group mean response rates during the second test phase. cs1 and cs2 occurred
on reinforced trials, when they were signalled by the occasion setter (OS: cs) and on
nonreinforced trials, when they were presented alone (cs). Error bars show within-
subject confidence intervals (Cousineau, 2005) calculated separately for each group.
For further details, see text and Table 9.
65