No evidence for olfactory blocking in honeybee classical conditioning

Over the last few decades, investigations into the role of reinforcement in the establishment of associations have remained central to associative learning research. One important discovery was that contiguous pairings of reinforcing and conditioned stimuli are not sufficient for associations to be formed. As argued recently by Fanselow (1998) and Holland (1993), the key experiment is the so-called ‘blocking’ experiment initially performed by Kamin (1968). In such an experiment, the first (pretraining) phase consists of pairings of stimulus A and reinforcement (A+) followed by a second (compound training) phase in which a stimulus B is added and the compound is reinforced (AB+). If stimulus B is tested later, relatively little response has developed to it despite pairing with reinforcement under otherwise optimal conditions. In particular, response levels to B are below those in control groups that have not received A+ pretraining. This suggests that pretraining with A+ is somehow capable of reducing (blocking) learning about B during AB+ trials. Thus, the presence of reinforcement during the AB+ trials is not sufficient to turn B into a conditioned excitor. Blocking was instrumental in the development of most current models of associative learning (Klopf, 1989; Mackintosh, 1975; Pearce and Hall, 1980; Rescorla and Wagner, 1972; Sutton and Barto, 1981; Wagner, 1981). The question of interest is why B does not turn into a conditioned excitor despite its contiguous pairing with reinforcement.

On a theoretical level, two classes of learning models have been suggested to explain blocking. As these models introduce concepts of either attention (Mackintosh, 1975; Pearce and Hall, 1980) or expectation and prediction (Rescorla and Wagner, 1972; Sutton and Barto, 1981), it has been argued that such cognition-like concepts must be introduced into the study of associative learning (e.g. Holland, 1993). The first class of model holds that blocking might come about by competition of conditioned stimuli for a limited attentional capacity (Mackintosh, 1975; Pearce and Hall, 1980), whereas the other proposes competition for the effects of reinforcement (Rescorla and Wagner, 1972; Sutton and Barto, 1981;). Thus, the former class of model holds that B does not enter into an association with the reinforcer because B itself is not effectively processed, whereas the latter holds that it is the reinforcer that is not effectively processed. On a physiological level, processing of either B or the reinforcer might thus be reduced to an extent that, at the relevant synapses, the internal signals for B and the reinforcer do not coincide during compound training, such coincidence being generally believed to be necessary for synaptic transmission to be altered associatively. In mammals, evidence exists both for alteration of conditioned-stimulus processing (Holland, 1997) and for modulation of reinforcement processing (Kim et al., 1998; Schultz, 1998).

As evidence for both modulation of reinforcement processing and for modulation of conditioned stimulus processing exists, one might suspect that the two kinds of modulation are not mutually exclusive. Thus, investigations into the role of reinforcement could benefit from a preparation in which blocking can be demonstrated behaviourally and in which a neurophysiological analysis of processing of both conditioned and reinforcing stimuli is possible. A classical conditioning paradigm in honeybees (Bitterman et al., 1983), in which harnessed honeybees learn to associate an odorant with a sucrose reward, seems to meet these demands (Hammer, 1997; Menzel and Müller, 1996). First, on a behavioural level, blocking between the elements of binary odorant compounds was first reported by Smith and Cobey (1994), and then analyzed further (Smith, 1996, 1997; Thorn and Smith, 1997). Smith (1996) suggested a role for attention in learning about odorants. Second, on a neural network level, an optical-imaging technique using Ca²⁺-sensitive and voltage-sensitive dyes was established that can be used to monitor odorant processing in vivo (Joerges et al., 1997; for methodological details, see Galizia et al., 1997) and that can identify associative modulations of olfactory processing (Faber et al., 1999). Third, on a cellular level, a single neuron has been identified that can serve the reinforcing function of a sucrose reward (Hammer, 1997) and that processes this reward in a flexible way: unpredicted rewards elicit higher spike frequencies than do predicted ones (Hammer et al., 1997).

The present study attempts to extend the behavioural analysis of blocking between the elements of binary odorant compounds in honeybee classical conditioning. We replicate a blocking experiment based on Smith and Cobey (1994) using the experimental design that was used most frequently in later studies (Smith, 1996, 1997; Thorn and Smith, 1997). In that design, pretraining consists of A+ trials in the BLOCK and of trials with a novel odorant C (C+) in the NOVEL control group. After both groups have received compound training with AB+ trials, lower test response levels to B in the BLOCK group are indicative of blocking. We will argue that this experiment is in itself not convincing evidence for blocking because we identify two confounding variables that have the potential to produce a blocking-like effect. First, according to Smith and Cobey (1994) (their Fig. 5), there was a decrease in the inter-trial interval (ITI) between pretraining and compound training (for details, see methods for experiment 1). There is ample evidence that the associative strength that a series of learning episodes can support increases with increasing ITI (Hintzman, 1974) (for details in honeybees, see Gerber et al., 1998). Therefore, following excitatory pretraining with A+, a decrease in ITI at the beginning of compound training could turn the added stimulus B into a conditioned inhibitor because it would signal a decrease in the associative strength supported by the learning episodes in the compound training phase. In a NOVEL control group, stimulus B would in any event become an excitor. Thus, a decrease in the ITI between pretraining and compound training would tend to decrease response levels to B in BLOCK compared with control groups, thus producing a blocking-like effect. Second, in the study of Smith and Cobey (1994), the experimental roles of the odorants were not balanced. For example, the novel odorant C was 1-octanol, whereas the use of geraniol and 1-hexanol was balanced as stimuli A and B. Thus, if the similarity between geraniol and 1-hexanol was less than that between either of these odorants and 1-octanol (or 2-octanone in Smith, 1996, 1997; Thorn and Smith, 1997), animals would generalize less from A to B in the BLOCK group than from C to B in the NOVEL control, again producing a blocking-like result.

Therefore, we performed further experiments of modified design to seek more convincing evidence for blocking. Our results will be discussed with respect to the persistent failure to demonstrate blocking between sensory modalities (for a review, see Bitterman, 1996; see also Couvillon et al., 1997; Funayama et al., 1995; Gerber and Smith, 1998). Furthermore, the problem of olfactory coding of binary compounds either by their elements or, alternatively, as novel unique perceptual entities will be discussed.

Exceptions to these general methods are described below along with each experiment; they were designed specifically to optimize procedures and/or make them similar to designs in previous studies on blocking from other researchers.

Honeybees (Apis mellifera L.) were caught in an indoor flight enclosure and harnessed as described in Bitterman et al. (1983). Harnessing allows free movement of only the mouthparts (including the proboscis) and antennae. Animals were then fed to satiation with a 1.25 mol l⁻¹ sucrose solution and kept overnight in a dark, cool (18–20 °C) and humid box. On the following day, the animals were tested for intact reflexes by touching one antenna with sucrose solution; if an animal showed an unconditioned response by completely extending its proboscis (as more than 90 % of animals do), it was used in the conditioning protocol. A complete response was scored if the proboscis was extended beyond the imaginary line between the opened mandibles. Conditioning began 30 min later by placing the animals on the training site. Olfactory stimuli and the sucrose reward were applied frontally; odorant-loaded air was removed by an exhaust system mounted behind the animals.

Trials lasted for 1 min. Soon (20 s) after the animals had been moved from their resting position to the training site, the sucrose-delivery syringe (precision syringe from Gilmont Instruments, Barrington, IL, USA) was placed approximately 1 cm behind the animal. After 6 s, the odorant was delivered for 4 s. Immediately (1 s) before odorant offset, a 1 μl droplet of 1.25 mol l⁻¹ sucrose solution was applied to both antennae and then offered for feeding, giving an onset–onset inter-stimulus interval (ISI) of 3 s and a 1 s overlap between odorant and sucrose reward. Reward delivery lasted for a total of 3 s.

Because in such trials the odorant precedes the reward, they are termed forward trials. Animals were then left untreated until a full minute had passed. Animals were returned to their resting position between trials, which were separated by an inter-trial interval of 10 min. Test trials involve presentation of the odorant only.

We used 1-hexanol, geraniol, 1-octanol and limonene as odorants. These four odorants can be applied in 24 different combinations, which were used in a balanced protocol. Below, we use ‘A, B, C, D’ to refer to a specific experimental role of a stimulus; letters do not refer to a particular chemical substance. All chemicals were obtained from Sigma.

We used three different methods of odorant application. For the first method (experiments 2, 6 and 7), 4 μl of odorant was loaded onto a strip of filter paper and placed into a 1 ml tuberculin plastic syringe, which was loaded into the olfactometer described in detail by Galizia et al. (1997). Binary compounds were produced by two separate syringes containing one filter paper each. An aquarium pump delivered a continuous air flow through the olfactometer, which had its opening at the training site 3 cm in front of the bee. Odorant pulses were applied by computer-controlled solenoid valves (Lee, Westbrook, CT, USA) programmed to shunt air through the respective odorant-loaded syringe. The valves were programmed such that they provided an unscented permanent air stream into which scented air could be injected.

The second method (experiment 1) was similar but used the less sophisticated olfactometer described by Smith and Cobey (1994), Smith (1996, 1997) and Thorn and Smith (1997).

In the third method (experiments 3–5 and 8), odorants were applied manually using 20 ml syringes as described in detail by Bitterman et al. (1983) and used in other investigations (e.g. Gerber et al., 1998; Müller, 1996), including studies reporting blocking (Smith and Cobey, 1994). Odorant syringes were freshly prepared daily.

The data from all animals used in experiments are reported. Only animals that failed to show an unconditioned response at the end of the experiment (<5 %) were operationally defined as ‘dead’ and were excluded from analysis.

During all trials, any complete proboscis extension after odorant onset and before reward onset was scored as a ‘response’. During backward trials (see experiment 3), a ‘response’ was scored if the proboscis was already extended at the time of reward onset; thus, performance on backward trials can give an estimate of the response-releasing properties of the ambient contextual stimuli. Data are presented as a population measure: the percentage of honeybees showing proboscis extension (%PE).

Within groups, response frequencies on single trials were compared using the McNemar test. χ²-tests were employed for corresponding between-group analyses.

Within groups, differences in response levels during the respective training and extinction phases were analysed using Wilcoxon signed-ranks tests. Between groups, Kruskal–Wallis tests were used for multiple-group comparisons of response levels during training as well as during extinction phases; for the corresponding two-group comparisons, Mann–Whitney U-tests were used.

Experiment 1 replicates an experimental design taken from the literature in which we have identified two procedural confoundations. First, Smith and Cobey (1994) (see their Fig. 5) introduced a decrease in the ITI between pretraining and compound training. Second, in Fig. 4 of Smith and Cobey (1994) and in Smith (1996, 1997) and Thorn and Smith (1997), the experimental roles of the odorants were not balanced.

Two groups of animals were trained: the BLOCK group was trained with A during six pretraining trials and with the AB compound during six compound training trials (A+/AB+), whereas the NOVEL control group was trained to a different stimulus C during pretraining, which was then not included during compound training (C+/AB+). Both groups were then tested once for their responses to B. Blocking would be indicated by lower response levels to B in the BLOCK than in the NOVEL group. A 2 mol l⁻¹ sucrose solution was used as the reward.

As in Smith and Cobey (1994) (their Fig. 4), the experimental roles of the odorants were not balanced: 1-hexanol and geraniol were balanced as A and B, but the novel stimulus C was always 1-octanol. In later studies (Smith, 1996, 1997; Thorn and Smith, 1997), 2-octanone was used as odorant C.

In keeping with Smith and Cobey (1994) (their Fig. 5), we introduced a decrease in the ITI between pretraining and compound training by pseudorandomly interspersing placement or ‘dummy’ trials during pretraining (A+, dummy, dummy, A+, dummy, A+, A+, dummy, A+, dummy, dummy, A+) but not during compound training (AB+, AB+, AB+, AB+, AB+, AB+). A ‘dummy’ trial involves placement to and from the experimental site only, but no explicit stimulation with either odorant or reward. Smith and Cobey (1994) introduced these dummy trials to keep the total number of trials during pretraining in the blocking group the same as in an additional unpaired control group in which pretraining consisted of presentations of either reward or odorant on separate trials and in a pseudorandom sequence. This unpaired control group thus had twice the number of trials compared with other groups so that dummy trials were added to the other groups to make up the difference. There was no difference in trial number between groups during compound training, and therefore no dummy trials were added in this phase. Thus, during compond training, the ITI between successive AB+ trials was always 10 min, whereas in the pretraining phase the ITI between successive A+ trials was variable and often longer. This experiment was conducted in the laboratory of Professor B. H. Smith, and therefore we could use the computer-controlled odorant-application device (providing a permanent air flow during periods when no odorant was applied) and the same general methods as reported by Smith and Cobey (1994), Smith (1996, 1997) and Thorn and Smith (1997).

Under the confounding conditions of experiment 1, evidence for a blocking-like effect was found (see below). Experiment 2 was designed to investigate whether a blocking-like effect can also be demonstrated using a within-animal design. In the pretraining phase, animals were trained with six rewarded trials to A. The compound training phase involved six rewarded training trials each to the compounds AB and CD in pseudorandomized order (e.g. AB, CD, CD, AB, CD, AB, AB, CD, AB, CD, CD, AB). For half the animals, the first trial in the compound training phase was an AB trial, for the remaining half it was a CD trial. At the regular 10 min ITI, animals were then tested for their responses to B and D (test). Lower response levels to B than to D would indicate blocking. Half the animals were tested with B first and the remaining half with D first, this sequence being balanced with respect to whether AB or CD was presented first during compound training.

Since preliminary experiments did not show evidence for blocking during the test, we introduced an additional, cumulative extinction phase on the next day in which seven unrewarded trials each with B and D were presented. The sequence of stimulus presentation was pseudorandom and balanced with respect to the two other sequences outlined above. The overnight rest period is predicted to increase response levels because it allows further memory consolidation and increases hunger motivation. This approach was used because Miller et al. (1996) showed that blocking might not be detectable immediately after training but might develop over time. All blocking experiments (except experiments 7 and 8, in which the extinction phase occurred immediately after training) include this next-day extinction phase. Sample sizes during the next-day extinction phase were reduced by mortality by approximately 10–20 %. However, the variable of primary interest for comparison with previous studies is the test immediately after the training. We therefore include the test performance of as many animals as possible, regardless of whether they survived to the next day.

Animals in this experiment remained in front of the exhaust system between trials. Furthermore, the olfactometer did not provide a continuous air flow during periods when no odorant was applied. This means that, during odorant application, the olfactory stimulus is accompanied by a mechanosensory (air puff) component; this mechanosensory component is, however of substantially lower salience than the olfactory one (Menzel, 1990; see also Pelz et al., 1997). The presence of a mechanosensory component resembles the procedure used by Smith and Cobey (1994).

Experiment 2 did not provide evidence for blocking (see below). Experiment 3 was designed to resemble the procedures employed by Smith and Cobey (1994) more closely and to overcome two procedural problems introduced by the design of experiment 2.

First, experiment 2 might have led to an increase in ITI between element pretraining and compound training. The ITI for the rewarded trials during element pretraining was 10 min. In the compound training phase, however, AB and CD trials were administered in pseudorandom sequence with 10 min between each successive trial. Thus, the pseudorandomization procedure leads to ITIs for each compound that are variable and often longer than 10 min. Therefore, since the reward during compound training occurs at unpredictable times, learning about B might remain intact. In experiment 3, element pretraining now consisted of both rewarded trials with A (seven forward trials, A+) and trials with C during which the sequence of the stimulus and reward was reversed, i.e. sucrose preceded odorant (seven backward trials, +C) with odorant stimulation starting 1 s before reward offset. This eliminates an ITI shift between element pretraining and compound training.

To determine whether blocking can be demonstrated using either constant or variable ITIs, two groups of animals were trained; one (the VARIABLE group) with a pseudorandom sequence of trial types. Ten minutes was allowed between each successive trial. The other group (the CONSTANT group) was trained with alternating A+ and +C as well as AB+ and CD+ trials. In keeping with Smith and Cobey (1994), 8 min 34 s (514 s) was allowed to elapse between successive trials in the CONSTANT group. In Smith and Cobey (1994), the ITI was ‘6 or 8 min’.

Second, the design of experiment 2 might have led to confounding effects of learned equivalence (Hall, 1996). In experiment 2, animals were trained using the scheme A+, AB+/CD+. Thus, animals might learn that any odorant will be rewarded and might actively broaden their generalization profiles. Since forward trials (A+) establish excitatory associative memories, whereas backward trials (+C) support no learning or even inhibitory learning (Hellstern et al., 1998), element pretraining with A+/+C is a discrimination procedure that can be expected to enhance differential processing of B and D during AB+/CD+ compound training. However, even in this design, the compound training phase (AB+/CD+) might introduce some level of learned equivalence (see below). Blocking would be indicated by lower response levels to B than to D during either the test or the seven extinction trials with each stimulus during the next-day extinction phase.

To make our procedures more similar to those used by Smith and Cobey (1994), we did not use the computer-controlled odorant delivery device of experiment 2 but applied the odorant manually using a 20 ml plastic syringe. This is a robust and standard procedure (Gerber et al., 1998; Müller, 1996; Smith and Cobey, 1994). In addition, in experiment 3 and in all following experiments, animals were removed from the experimental site between trials (as is also a standard procedure). The use of geraniol, 1-hexanol, 1-octanol and limonene as odorants was completely balanced.

Experiments 2 and 3 did not provide evidence for blocking (see below). Experiments 2 and 3 used two different within-animal designs. Experiment 4 was designed to use a more conventional (Smith and Cobey, 1994) between-animal design. In within-animal designs, learned equivalence might be established during compound training (Hall, 1996), which could obscure blocking. This is because individual animals receive rewarded trials with both AB and CD. Such training may result in actively broadened generalization and thus obscure possible differences in response levels to B and D. A between-animal design can reduce such equivalence effects, at least during the compound training phase.

Two groups of animals were trained with differences in their experience during pretraining. The BLOCK group received seven forward trials with A, whereas the CONTROL group received seven backward trials with A. Both groups were then trained seven times to AB and were tested once for responses to B; on the next day, animals were again tested for their responses to B during the seven trials of the next-day extinction phase. Lower response levels to B in the BLOCK than in the CONTROL group would indicate blocking. This procedure closely resembles that of Fig. 5 in Smith and Cobey (1994), although the latter study did not include a next-day extinction phase. To follow their procedure as well as that of experiment 3, the ITI in experiment 4 was 8 min 34 s (514 s). The use of geraniol, 1-hexanol, 1-octanol and limonene as odorants was completely balanced.

Experiments 2–4 did not provide evidence for blocking (see below). To conform with previous studies on blocking (Smith and Cobey, 1994), in experiment 5 we used a more restricted set of odorants. Experiment 4 was repeated but with only two odorants: 1-hexanol and geraniol used in a balanced manner. We ensured that a sufficient sample size was obtained for both combinations of odorants; for the backward pairing control relevant at this point, Smith and Cobey (1994) used only one combination of odorants: 1-hexanol as A and geraniol as B.

Experiments 2–5 did not provide evidence for blocking (see below). Experiment 6 was designed to test the effects of the method of odorant application on the results of experiment 5. Experiment 6 used the computer-controlled odorant-delivery device also used in experiment 2 (see Galizia et al., 1997). Unlike in experiment 2, however, the computer was programmed to deliver a continuous clean air stream into which odorant could be injected. This removes the mechanosensory component of odorant application (Pelz et al., 1997) that was present in experiments 1–5. The presence of stimuli additional to odorants A and B could theoretically (e.g. Rescorla and Wagner, 1972) reduce any blocking effect. This is because the share in associative strength that the compound’s elements gain from training decreases as the number of elements involved increases, thus potentially leading to a ‘floor’ effect.

Experiments 2–6 did not provide evidence for blocking (see below). Experiment 7 was designed to test whether responses to B in the BLOCK group are indeed due to associative memories for B. This is important because it is possible that responses to B in the BLOCK group are generalization responses, whereas in the CONTROL group the responses to B might be based directly on memories for B. That there may be stronger generalization in the BLOCK than in the CONTROL group seems likely because, over the two phases of the experiment, the BLOCK group essentially receives an equivalence training (A+, AB+) whereas the CONTROL group receives differential conditioning (+A, AB+).

Four groups of animals were trained, two of which were BLOCK groups and two of which were CONTROL groups; groups were then tested for their responses either to B (BLOCK^B and CONTROL^B) or to a novel stimulus C (BLOCK^GEN and CONTROL^GEN). Lower response levels to B in the BLOCK^B than in the CONTROL^B group would indicate blocking. A learned equivalence effect would be indicated by higher response levels to C in the BLOCK^GEN than in the CONTROL^GEN group.

We used four odorants: geraniol, 1-hexanol, 1-octanol and limonene in a balanced way. Since in experiments 1–6 the next-day extinction phase did not yield additional information (see below), we ran the extinction phase immediately after training. All other variables were as in experiment 6.

Experiments 2–7 did not provide evidence for blocking (see below). In the initial reports of blocking (Kamin, 1968), a backward pairing control was not used, but instead results were compared with a control group which did not receive any pretraining. For the honeybee, Smith and Cobey (1994) introduced the use of a number of additional control procedures, including the backward control; from those controls, subsequent studies (Smith, 1996, 1997; Thorn and Smith, 1997) used the NOVEL control group, which received forward pairings with a novel odorant C during element pretraining. Experiment 8 was designed to investigate whether a blocking-like effect could be demonstrated using four different control groups that received pretraining with (1) backward pairings of A, or (2) presentation of only the reward, or (3) of only the odorant, or (4) forward pairings with a novel stimulus C. This last NOVEL control is of particular interest because it is the control group used most frequently in the existing honeybee literature; furthermore, it does not suffer from the confounding variables in design that we suggested above could be responsible the blocking-like effect reported in experiment 1 of our study and by Smith and Cobey (1994), Smith (1996, 1997) and Thorn and Smith (1997) (see Introduction). Thus, comparing the outcomes of experiments 1 and 8 of the present study with respect to the NOVEL control can give an estimate of how powerful those confounding variables might be in supporting a blocking-like effect.

Five groups of animals were trained. All groups received compound training with AB and were tested for B. Groups differed with respect to pretraining: they received forward pairings of A and reward (BLOCK group), backward pairings of A (CONTROL group), forward pairings with a novel odorant C (NOVEL group), odorant A only (ODORANT group) or reward only (REWARD group). Any difference in response levels towards B would indicate a blocking-like effect; the nature of this effect would vary according to the groups between which differences were found. The ITI was 10 min. All other variables were as in experiment 7.

Response levels to B during the test were significantly lower in the BLOCK than in the NOVEL group (Fig. 1, P<0.05, χ²=11.31, d.f.=1), indicating the existence of a blocking-like effect. Between-group differences were not found during compound training (Fig. 1, P>0.05, U=173.5) or pretraining (Fig. 1, P>0.05, U=175.5). Interestingly, the between-group comparison during compound training gives an estimate of perceived stimulus similarity: as the NOVEL group was trained with C+ during pretraining, response levels to AB by this group at the beginning of compound training partially reflect the perceived similarity between AB and C. This phenomenon of responses to stimuli that have themselves not been trained but which are similar to trained responses is termed stimulus generalization. Importantly, although pretraining with A was not sufficient to increase response levels to AB above generalization levels (Fig. 1, compound training), it was sufficient to induce a blocking-like effect (Fig. 1, test). This effect therefore occurs despite a lack of evidence that animals have learned specifically about the blocking stimulus A, confirming that variables different from or additional to the associative strength accrued to A might be involved in producing a blocking-like effect.

Experiment 2 did not provide evidence for blocking in either the test or the next-day extinction phase; in both instances,response levels to B versus D did not differ significantly (Fig. 2; test, P>0.05, χ²=0.03, d.f.=1; next-day extinction phase, P>0.05, W=585).

In the compound training phase, animals showed significantly higher response levels to compound AB than to CD (Fig. 2, P<0.001, W=930.5), indicating that excitatory associative memories for A are the basis for the conditioned response to AB in the compound training phase. At the very least, this implies that animals generalize more strongly from A to AB than from A to CD. This suggests that an AB compound does have element-A-like properties and also that evidence for blocking is lacking even though there is evidence that animals have learned specifically about the blocking stimulus A. Together with the results of experiment 1, this provides a double dissociation between the occurrence of a blocking-like effect and the associative strength of the blocking stimulus A. Furthermore, all the following experiments also show evidence for specific memories for A in the apparent absence of a blocking-like effect (see below).

Experiment 3 also did not provide evidence for blocking. Response levels to B and D were not statistically different in both the test (Fig. 3; VARIABLE, P>0.05, χ²=0.31, d.f.=1; CONSTANT, P>0.05, χ²=0.64, d.f.=1) and the next-day extinction phase (Fig. 3; CONSTANT, P>0.05, W=188; VARIABLE, P>0.05, W=142).

During compound training, both groups showed significantly higher response levels to compound AB than to CD (Fig. 3; VARIABLE, P<0.001, W=9; CONSTANT, P<0.001, W=1), indicating that, as in experiment 2, the basis for responses to AB were excitatory associative memories for odorant A resulting from element pretraining. Thus, the compounds have element-like properties.

A response was scored whenever the proboscis was extended before application of the sucrose reward. On forward but not backward pretraining trials, sucrose application was preceded by odorant stimulation, so that comparing response levels between forward and backward trials will give an estimate of the extent to which contextual stimuli (handling, visual, mechanosensory, etc.) can control responses. In both groups, response levels on pretraining forward trials were much higher than on backward trials (Fig. 3; VARIABLE, P<0.001, W=0; CONSTANT, P<0.001, W=0), suggesting that contextual stimuli alone barely release responses.

No evidence for blocking was found in experiment 4. Reponse levels to B in the BLOCK group were not statistically different from those in the CONTROL group (Fig. 4A) in both the test (P>0.05, χ²=0.35., d.f.=1) and the next-day extinction phase (P>0.05, U=666).

In the compound training phase, response levels to AB in the CONTROL group were significantly lower than in the BLOCK group (Fig. 4A, P<0.001, U=404.5), suggesting that pretraining with A+ results in excitatory associative strength in the BLOCK group that supports responding to AB. Thus, as in the previously described within-animal designs, the AB compound is likely to have A-like properties.

During pretraining, response levels in the BLOCK group were significantly higher than in the CONTROL group (Fig. 4A, P<0.001, U=30), confirming that the responses observed in the BLOCK group result from odorant stimulation rather than contextual stimuli (see above).

Experiment 5 also did not provide evidence for blocking; the results are identical to those of experiment 4. There was no significant difference in response levels to B between the BLOCK and CONTROL groups (Fig. 4B; test, P>0.05, χ²=0.08, d.f.=1; next-day extinction phase, P>0.05, U=855.0). If the data are separated by odorant (Fig. 4C), the same pattern is found for both the test (1-hexanol as B, P>0.05, χ²=0.16., d.f.=1; geraniol as B, P>0.05, χ²<0.01., d.f.=1) and the next-day extinction phase (1-hexanol as B, P>0.05, U=182; geraniol as B, P>0.05, U=240.5).

Response levels to AB in the CONTROL group during compound training were significantly lower than in the BLOCK group (Fig. 4B, P<0.001, U=764.0), showing that excitatory associative strength for A has developed in the BLOCK group that supports responding to the AB compound and that, as in experiments 2–4, AB has A-like properties.

During pretraining, response levels in the BLOCK group were significantly higher than in the CONTROL group (Fig. 4B, P<0.001, U=93.5), showing that responses in the BLOCK group were due to odorant stimulation and not contextual stimuli (see above).

Experiment 6 also did not provide evidence for blocking; the results are identical to those from experiments 4 and 5. Response levels to B in the BLOCK and CONTROL groups were statistically indistinguishable (Fig. 4D) in both the test (P>0.05, χ²=0.14., d.f.=1) and next-day extinction phase (P>0.05, U=1067.5). The same pattern was found when the data were separated by odorant (Fig. 4E) in both the test (1-hexanol as B, P>0.05, χ²<0.01., d.f.=1; geraniol as B, P>0.05, χ²=0.31., d.f.=1) and next-day extinction phase (1-hexanol as B, P>0.05, U=272.0; geraniol as B, P>0.05, U=274.0).

During compound training, response levels to AB were significantly lower in the CONTROL than in the BLOCK group (Fig. 4D, P<0.05, U=1527.0), showing that pretraining results in excitatory associative strength in the BLOCK group that supports responding to AB, again demonstrating the A-like properties of the AB compound (see above).

During pretraining, response levels in the BLOCK group were significantly higher than in the CONTROL group (Fig. 4D, P<0.001, U=729.5), showing that responses in the BLOCK group were due to odorant stimulation rather than to contextual stimuli (see above).

Experiment 7 did not provide evidence for either blocking or learned equivalence. Across groups, response levels during extinction differed significantly (Fig. 5, P<0.05, H=9.95, d.f.=3). Specifically, a comparison between the two BLOCK groups showed higher response levels to B than to C (Fig. 5; BLOCK^Bversus BLOCK^GEN, P<0.05, U=1663.0); the same relationship was found between the two CONTROL groups (Fig. 5; CONTROL^Bversus CONTROL^GEN, P<0.05, U=1795.5), indicating that, regardless of forward or backward pretraining with A, response levels to B were above generalization levels. Response levels to B in the BLOCK^B group were not statistically different from those in the CONTROL^B group (Fig. 5, P>0.05, U=1990.0), indicating that blocking was not evident. Response levels to C in the BLOCK^GEN group did not differ significantly from those in the CONTROL^GEN group (Fig. 5, P>0.05, U=2336.0), indicating that learned equivalence was not evident.

During compound training, response levels to AB were significantly higher in the BLOCK than in the corresponding CONTROL groups (Fig. 5; BLOCK^Bversus CONTROL^B, P<0.01, U=1036.5; BLOCK^GENversus CONTROL^GEN, P<0.01, U=1471.5), suggesting that pretraining results in excitatory associative strength in the two BLOCK groups that supports responses to AB, indicating that compound AB has A-like properties (see above). Neither the two BLOCK groups nor the two CONTROL groups differed significantly during compound training (Fig. 5; BLOCK^Bversus BLOCK^GEN, P>0.05, U=2117.5; CONTROL^Bversus CONTROL^GEN, P>0.05, U=1986.0), making it unlikely that the differences observed were due to spurious differences in group composition.

During pretraining, response levels in the two BLOCK groups were significantly higher than in the corresponding two CONTROL groups (Fig. 5; BLOCK^Bversus CONTROL^B, P<0.001, U=264.5; BLOCK^GENversus CONTROL^GEN, P<0.01, U=444.0), showing that responses in the BLOCK groups were due to odorant stimulation and not to contextual stimuli (see above). Neither the two BLOCK groups nor the two CONTROL groups differed significantly during pretraining (Fig. 5; BLOCK^Bversus BLOCK^GEN, P>0.05, U=2169.0; CONTROL^Bversus CONTROL^GEN, P>0.05, U=2225.0), making it unlikely that the differences observed were due to spurious differences in group composition.

Response levels to B differed significantly between groups (Fig. 6) both during the test (P<0.05, χ²=11.50, d.f.=4) and during the complete extinction phase (P<0.05, H=10.65, d.f.=4). Specifically, response levels in the ODORANT group were higher than for the other groups (Fig. 6; REWARD versus ODORANT, P<0.05, U=3501.0; CONTROL versus ODORANT, P<0.001, U=3595.0; BLOCK versus ODORANT, P<0.01, U=3231.0), and the same trend was found for the NOVEL group, although this did not quite reach statistical significance (Fig. 6, P=0.06, U=3456.0). All other pairwise comparisons for the complete extinction data for B were not statistically significant; in particular, as in experiments 4–7, the BLOCK and CONTROL groups did not differ (Fig. 6, P>0.05, U=4138.0). This shows that the presence of a blocking-like effect is not specific for the predictive value of the pretrained stimulus A. Furthermore, the lack of a significant difference between the BLOCK and NOVEL groups (Fig. 6, P>0.05, U=4027) is in contrast to the results of Smith and Cobey (1994), Smith (1996, 1997) and Thorn and Smith (1997) and to our results from experiment 1. Experiment 8 was designed to avoid the confounding effects of asymmetry in odorant use present in the above studies and in experiment 1. Under these unconfounded conditions, there is no evidence for a blocking-like effect using a NOVEL control.

Experiment 8 shows that, regardless of forward or backward pairing of A, of training with a novel odorant C or of mere exposure to reward alone, the same amount of learning accrues to a supposedly blocked stimulus B. The only significant difference was for animals that did not receive any reward during pretraining (the ODORANT group). The most plausible explanation for this is that animals in the ODORANT group were hungrier than those from the other groups. This does not mean that the animals in the other groups are satiated and not sufficiently motivated to show responses: for these groups, conditioned response levels had attained uniformly (Fig. 6, P>0.05, χ²=7.55, d.f.=3) high levels of approximately 70 % at the end of compound training. Furthermore, our analysis includes only those animals that showed intact unconditioned responses immediately after the last extinction trial. Thus, the effects of satiety, if present, are not large enough to reduce response levels to the extent that would render a putative blocking effect undetectable. An alternative explanation for the significantly higher response levels in the ODORANT group could be that all the other groups became habituated during pretraining. However, the levels of conditioned responses during pretraining and compound training seem to be too high and too stable for this to be a plausible conclusion. Furthermore, it has been shown that approximately 200 trials at ITIs of 15 s are needed to cause habituation (Braun and Bicker, 1992), making it unlikely that seven trials at a 10 min ITI will have the same effect.

If response levels to AB depend on pretraining of A such that forward pairings of A lead to the highest response levels to AB and backward pairings to the lowest, it could be that forward pairings establish excitatory, but backward pairings inhibitory, associative strength. In this context, experiment 8 might be viewed as testing the extent to which different pretraining procedures with A can affect acquisition of AB. Indeed, response levels during compound training differed significantly between groups (Fig. 6, P<0.001, H=55.67, d.f.=4); specifically, response levels in the BLOCK group were significantly higher than in any other group (Fig. 6; BLOCK versus CONTROL group, P<0.01, U=1999.0; BLOCK versus NOVEL group, P<0.01, U=3364.0; BLOCK versus REWARD group, P<0.01, U=2684; BLOCK versus ODORANT group, P<0.01, U=2656.0), indicating that in the BLOCK group responses to AB are due to excitatory associative memories for A. Response levels during compound training in the CONTROL group were significantly lower than in any other group (Fig. 6; CONTROL versus NOVEL group, P<0.01, U=2734.0; CONTROL versus REWARD group, P<0.01, U=3480.0; CONTROL versus ODORANT group, P<0.01, U=2598.0), indicating that in the CONTROL group response levels to AB might be reduced by inhibitory associative memories for A. Thus, forward pairing of A and reward leads to excitatory learning, whereas backward pairing seems to lead to inhibitory learning. This latter result is to some extent unexpected because Hellstern et al. (1998) showed that backward ISIs of 15 s led to maximal inhibitory learning, whereas either shorter (6 s) or longer (30 s) ISIs did not yield inhibition. However, Hellstern et al. (1998) used only three or fewer conditioning trials; it is possible that more trials (at least seven) are needed before significant effects can be observed for ISIs as short as the 3 s used in the present study (Wagner and Terry, 1975). An alternative explanation for our results might simply be that, in the CONTROL group, no excitation (rather than inhibition) developed towards A. In this case, the lower response levels of the CONTROL group compared with the REWARD and ODORANT groups to AB would imply that the animals in the latter two groups did learn about A in an excitatory way during pretraining. This seems unlikely, since Bitterman et al. (1983) have shown that odorant pre-exposure has inhibitory effects. We believe that the most plausible explanation of our results is that backward pairings of odorant and sucrose, even with these short ISIs, establish inhibitory learning. In any event, if the backward pairing ISIs used in experiments 4–8 did induce inhibitory associative strength, such inhibition would increase the differential processing of B during AB compound training in BLOCK versus CONTROL groups. That is, since excitatory learning of A is predicted to block B, inhibitory learning should enhance learning of B, thus leading to particulary pronounced blocking, if blocking were to occur.

As noted above, response levels to AB are higher in the BLOCK than in the NOVEL group, providing evidence for specific memories for A in the BLOCK group. This contrasts with the result of experiment 1, in which no difference in response levels to AB was found. This provides a double dissociation between a blocking-like effect (evident in experiment 1 but not experiment 8) and specific memories for A (evident in experiment 8 but not experiment 1).

Finally, performance during pretraining also differed significantly between groups (Fig. 6, P<0.001, H=264.07, d.f.=4); but the performance of the BLOCK and NOVEL groups did not differ (Fig. 6, P>0.05, U=3919.0), showing that the observed differences were not due to spurious differences in group composition.

By exploiting the proboscis extension reflex (PER), the present study examined the presence of blocking in honeybees using odour–odour binary compound stimuli. Regardless of experimental design, time of testing, the number of extinction trials, the method of odorant application or odorant identity, blocking could not demonstrated to be present (experiments 2–6). Response levels to the test stimulus were equal and equally above generalization levels in BLOCK and backward CONTROL groups (experiment 7). Response levels in the BLOCK group were lower than in an ODORANT control, probably due to differences in hunger motivation (experiment 8). If the experiments are performed under the confounding conditions of a decrease in the ITI between pretraining and compound training and with the odorants having asymmetric experimental roles, a blocking-like effect may be detectable relative to a NOVEL control (experiment 1). Under unconfounded conditions, however, this effect cannot be demonstrated (experiment 8).

The only report of blocking using a backward pairing control is in Fig. 5 of Smith and Cobey (1994). Specifically, they reported that response levels to geraniol (stimulus B) differ depending on whether geraniol was trained in compound with 1-hexanol (stimulus A) that had been either pretrained in a forward (BLOCK, A+/AB+/test B) or backward (CONTROL, +A/AB+/test B) fashion. Response levels to B were lower after forward than after backward pretraining of A, a result taken to indicate blocking. Our attempts to replicate this result failed (Fig. 4C,E, right-hand parts). Attempts to find an effect in experiments which balanced the experimental roles of geraniol and 1-hexanol also failed (Fig. 4B,D), as did experiments using a wider range of odorants (Figs 4A, 5, 6) or that used a within-rather than between-animal design (Fig. 3). Concerning the backward pairing control, we are not aware of parametric deviations from Smith and Cobey (1994) that could have diminished blocking. However, concerning the study by Smith and Cobey (1994), three procedural arguments might be made.

• (1) There was a decrease in the ITI between pretraining and compound training that in the BLOCK group but not in the CONTROL could have turned B into a conditioned inhibitor. Thus, response levels to B would be lower in the BLOCK group compared with the CONTROL (see Introduction for details). Comparison of the NOVEL control versus BLOCK group values of experiment 1 (Fig. 1) and experiment 8 (Fig. 6) indeed suggests that a decrease in ITI might be a critical variable for the detectability of a blocking-like effect.

• (2) The experimental roles of the stimuli were not balanced. Therefore, asymmetries in generalization may be sufficient to explain a blocking-like effect as reported by Smith and Cobey (1994) (see Introduction for details).

• (3) Smith and Cobey (1994) used the following criterion to determine which animals were included in their statistical analysis: ‘Any subject that showed no response to the associated odorant during acquisition trials and/extinction (sic) tests was not used in statistical analyses’. In the present study, the only subjects that were excluded were those defined as ‘dead’ (see General methods). Selection criteria can be justified only as long as they do not introduce asymmetries between groups, but just such an asymmetry may have been introduced by Smith and Cobey (1994). In their BLOCK group, some subjects might have been excluded because they did not show a conditioned response during pretraining. In the backward pairing CONTROL, however, this criterion cannot be applied, at least not during pretraining.

Taken together, we believe that these procedural problems in Smith and Cobey (1994) should raise substantial concerns. Concerning the BACKWARD control group, they seem to be capable of explaining the discrepancies between the results of Smith and Cobey (1994) and our results, which excluded these potentially confounding effects. We therefore believe that at present there is no convincing evidence in honeybee classical conditioning for the existence of a blocking effect that is specific for the predictive nature of the blocking stimulus.

After the initial report of blocking (Smith and Cobey, 1994), several follow-up studies were published (Smith, 1996, 1997; Thorn and Smith, 1997) that restricted their investigations to the blocking procedure A+/AB+/test B using as control a group that received pretraining with a novel odorant (C+/AB+/test B) or a group that received no pretraining at all. Two questions need to be addressed.

• (1) What is the analytical value of a novel or no-pretraining versus a backward control? We believe that comparisons with a backward control allow more specific conclusions to be drawn. If a novel control is used, the effects cannot be distinguished from those of odorant pre-exposure. To resolve this, an additional odorant-only control (A/AB+/test B) should be added. However, this control would not be equivalent regarding the total number of reward stimulations. The same arguments apply to the use of a no-pretraining control. Indeed, experiment 8 showed that omitting a reward during pretraining can produce a blocking-like result that is probably caused by a motivational process. Smith and Cobey (1994) addressed this problem with the use of an unpaired control (A↔+/AB+/test B); however, as noted above, this introduced a decrease in the ITI between phases and thus introduced another potentially confounding variable. The use of a backward control, however, avoids ITI shifts between phases and equates exposure to the blocking odorant, total reward delivery and temporal overlap between odorant and reward; the only difference between the groups in this case being the predictive value of the pretraining odorant.

• (2) Regardless of its heuristic value, why could experiment 8 not replicate the outcome for the novel control as reported by Smith and Cobey (1994) (their Fig. 4), Smith (1996, 1997) and Thorn and Smith (1997) It should be noted that experiments in those studies do not suffer from the potentially confounding effects of either post-hoc selection criteria or of ITI shifts between phases. However, the present experiments, but not those in the cited studies, were balanced for the experimental roles of the odorants. In experiment 8, all 24 possible combinations of 1-hexanol, geraniol, 1-octanol and limonene were used equally often. The studies cited above balanced the use of 1-hexanol and geraniol as stimuli A and B, respectively; however, the novel odorant was always 1-octanol (Smith and Cobey, 1994) (their Fig. 4) or 2-octanone (Smith, 1996, 1997; Thorn and Smith, 1997). Therefore, regarding the novel control, all previously reported cases of blocking were obtained with asymmetries in the experimental roles of the stimuli involved. The blocking-like effect found in those experiments might therefore have been caused by a weaker similarity between geraniol and 1-hexanol than between either of these odorants and 1-octanol (or 2-octanone). In consequence, animals would generalize less from A to B in the blocking group than from C to B in the novel control. Indeed, in our experiment 1, which incorporated similar asymmetries in odorant use, a blocking-like effect was detectable that is not evident under unconfounded conditions (experiment 8), indicating that such asymmetries may be critical for the detectability of a blocking-like effect.

Most current models of conditioning clearly predict blocking (Klopf, 1989; Mackintosh, 1975; Pearce and Hall, 1980; Rescorla and Wagner, 1972; Sutton and Barto, 1981). Thus, the apparent absence of blocking in honeybee classical conditioning seems to be at odds with mainstream concepts developed for vertebrates. A possible reason for this discrepancy could be that all these models assume that the elements of compounds retain their elemental integrity (i.e. they are elementary models). There is now an increasing volume of evidence that olfactory compounds are largely perceived as unique, novel stimuli and that the ability of subjects to perceive a compound by its elements is very limited (Laing and Francis, 1989; Laurent, 1996; Livermore et al., 1997). Therefore, a compound made up from the elements A and B would largely be perceived as a new configuration X. Importantly, X is thought to have a stronger similarity to its elements than to a novel stimulus such as C; to illustrate this assumption, we will write X as X^ab. This is a concept akin to the configural learning model of Pearce (1994), which proposes that it is not the elements A and B of a compound AB but instead a new configural unit X^ab that enters into associations with reinforcement. Regarding compound processing in the present study, two arguments suggest that olfactory compounds are indeed processed in this way. First, honeybees can solve a biconditional discrimination task of the form AB+, CD+/AC−, BD−, which would not be possible using elementary processing but is easily possible using configural processing (Chandra and Smith, 1998; Hellstern et al., 1995). Second, an analysis of sensory preconditioning has shown that configural models can account for the observed pattern of results much better than can elementary ones (Müller et al., 1996). Furthermore, results on single trial overshadowing (Pelz et al., 1997; Smith, 1996) can be accounted for by configural models. However, an experiment by Hellstern et al. (1998) showed that the excitatory associative strength established on forward trials summates with the inhibitory associative strength established on backward trials; this might suggest that, under some experimental conditions, honeybees can process a compound by its elements. Thus, it is possible that the extent to which honeybees process a compound in an either elementary or configural way depends on the demands imposed on the animals by the training conditions (see Livermore et al., 1997). Taken together, most results to date can be explained by configural processing, with the exception of the experiment by Hellstern et al. (1998), which suggests a role for elementary processing.

In the present study, response levels to AB at the beginning of compound training were as high as response levels to A at the end of pretraining, implying that this compound does have element-like properties. Elementary theory would argue that element A is detected in the AB compound and is responded to, whereas configural theory would argue that (at least at the beginning of compound training) the similarity between A and X^ab is high. Thus, this result is equally compatible with either theory and, importantly, blocking would be predicted by either theory under these conditions.

As outlined above, we believe that a configural model captures the nature of odorant mixture processing much better than an elementary theory. Therefore, we explain below how configural theory deals with blocking. We then suggest a modification of configural theory that has the potential to resolve the discrepancy between a high similarity between A and X^ab and the apparent absence of blocking.

One configural learning model of Pearce (1994) deals with blocking by incorporating a modified form of the competitive delta rule. The change in associative strength of a configural unit X^ab (E_X^ab) is proportional to the difference between the maximum associative strength a reinforcer can support (λ) minus the activation level of a ‘reinforcement unit’ (V_X^ab). V_X^ab is determined by summing the associative strengths of configural unit X^ab and the associative strengths of those units that are concurrently activated because of their similarity with X^ab; importantly, the weight of these additional units can be reduced by a similarity factor S (for details, see Pearce, 1987, 1994), which can range from 0 to 1. In a blocking experiment, pretraining is designed to increase the associative strength of A (E_A) to l. During compound training, the final level of associative strength of the compound X^ab (E_X^ab) would be limited by the fact that the reinforcement unit will already be activated by the compound according to its similarity to A (S ×E_A). Therefore, some learning about X^ab would occur, and test responses to B would be based on the similarity between B and X^ab and A. Most importantly, however, because the model incorporates a competitive delta rule, the associative strength accrued to X^ab would in all cases be inversely related to the associative strength that A had already gained during pretraining. Thus, the model of Pearce (1994) predicts blocking.

Compared with elementary theories, both classes of model thus predict that the more pronounced the associative effects of pretraining with A upon response levels to AB during compound training, the stronger the blocking effect should be. In this study, however, we provided a double dissociation between these two effects: in experiment 1, there is a blocking-like effect in the apparent absence of associative memories for A, whereas in experiments 2–8, there is evidence for associative learning about A in the apparent absence of blocking.

According to our results and conforming to the above configural learning model, we have to assume a strong similarity between X^ab and A. However, according to Pearce (1994), such a strong similarity would support strong blocking, which was not found in the present study. In its original form, the model of Pearce (1994) does therefore not adequately predict our results. Nevertheless, the model is attractive because it captures the largely configural nature of olfactory processing. Within this framework, the problem is therefore to reconcile the high similarity between X^ab and A with the apparent absence of blocking. This can be done by postulating that the similarity (S) between X^ab and A decreases with repeated compound presentations. This would result in a gradual decrease in the contribution of the pretrained element (S ×E_A) and would allow E_X^ab to increase to an asymptotic value.

There are some data on honeybee classical conditioning with odorant compounds that can be interpreted along these lines: Müller et al. (1996) found that a single compound exposure is sufficient to support sensory preconditioning. That is, animals were exposed to AB on a single occasion and were then tested for responses to B; in a second phase, A was reinforced and finally B was tested again. Response levels to B increased between the first and the second test. Müller et al. (1996) suggested that a single exposure to AB initiates a learning process that can establish a new configural unit X^ab. Because of its similarity to A, this configural unit gains associative strength during reinforcement of A; in the second test, B, because of its similarity with X^ab, releases responses that are ultimately based on associations of X^ab with reward. Importantly, this sensory preconditioning effect disappears with repeated compound presentations. If sensory preconditioning is indeed based on the similarity between X^ab and A, this could imply that repeated compound presentations decrease the perceived similarity between compounds and their elements. This modification to the model of Pearce (1994) could be tested by determining whether sensory preconditioning is weakened by repeated reinforced trials with A, whether overshadowing becomes weaker with repeated trials or whether pre-exposure treatments modulate the ease with which feature-positive or feature-negative discrimination tasks can be solved.

Alternatively, within the model of Pearce (1994), it could be suggested that those processes that establish configural units and those processes that release responses via these configural units are independent of the competitive aspects of the delta rule. In other words, the delta rule might apply and conditioning of B might indeed be blocked, but this phenomenon might be masked by a superimposed process that stores the conjunction of A and B into the configural unit X^ab and which, upon stimulation with B, releases responses on the basis of this conjunction. Such a process would mask blocking by decreasing response levels in a backward control in which B was presented in conjunction with the conditioned inhibitor A and by increasing response levels to B in the block group in which B was presented together with the conditioned excitor A. A way to test this model would be to extinguish the conjunction of A and B by unreinforced presentations of A after compound training.

A third possibility is that the delta rule itself, which is incorporated in both elementary and configural theory, might not be appropriate to the present study. All the theories mentioned above assume that the increase in associative strength of a stimulus is dependent on the associative strengths of all other stimuli present during that trial. On the basis of the absence of blocking between sensory modalities observed in a large series of experiments (see below), Bitterman (1996) suggested that this dependence might not exist in invertebrates, at least not for inter-modal compounds. This implies that the elements of a compound gain and lose associative strength independently of each other; furthermore, it implies a strong role for elementary processing. Although devised for inter-modal compounds, our results showing an absence of blocking in intra-modal compounds is clearly compatible with this independence assumption. However, the suggested strong role of elementary processing does not seem appropriate for olfactory processing (see above), and therefore we believe that the independence assumption does not apply to binary odorant compounds.

Blocking in classical conditioning of the proboscis extension reflex has also been investigated for inter-modal compounds. Gerber and Smith (1998) found no evidence for blocking between a pretrained visual stimulus A and an added olfactory stimulus B. On the contrary, BLOCK groups showed higher response levels to B than did a backward CONTROL. Thus, visual stimuli seem to enhance, rather than to block, olfactory learning. Importantly, this result refutes the independence assumption, at least for visual–olfactory compounds in PER conditioning.

Inter-modal blocking was also investigated in freely flying honeybees. A large number of experiments failed to detect blocking between various sensory modalities (for a review, see Bitterman, 1996). Recently, these studies have been confirmed (Funayama et al., 1995) and extended (Couvillon et al., 1997): no blocking was found between visual and olfactory stimuli; however, if intra-modal compounds were used in either vision or olfaction, blocking was detected. The experimental design used by Couvillon et al. (1997) was a ‘concurrent blocking design’. In this one-phase design, animals receive either of two trial types in a pseudorandom sequence: in eight out of 16 trials, AB is rewarded; in the remaining eight trials, animals receive an A versus C discrimination (not choice!) training with a 2 min ITI. In the block group, A is rewarded and C is not, whereas in the control group C is rewarded and A is not. Blocking would be indicated by lower response levels to B in the block versus the control group. That is, the experiment has the logical form AB+ (A+/C−), test B versus AB+ (A−/C+), test B. In principle, such a design could also be used for PER conditioning. However, it is likely to be less powerful than that used in the present study because, with our procedures, learning about A has already reached an asymptote at the beginning of compound training, whereas in a ‘concurrent blocking design’ the capacity of A to block B is initially low. In any event, an analogous design in PER conditioning would have to incorporate the following modifications. First, A versus C discrimination training would have to be carried out on separate trials. Second, the ITI during A versus C training should be long enough to allow for memory consolidation (>7 min; see Menzel, 1990; Gerber et al., 1998). Third, and possibly most importantly, one would have to use odors in all experimental roles, whereas Couvillon et al. (1997) used odors as stimuli A and B, but a visual stimulus as C. In PER conditioning, olfactory stimuli are known to have a much higher salience than visual stimuli (Gerber and Smith, 1998), whereas they have approximately equal salience in paradigms using freely flying bees. Future studies should attempt to identify isomorphic procedures for freely flying and harnessed honeybees, especially given the apparently conflicting evidence regarding olfactory blocking between freely flying and harnessed honeybees. However, there is good reason to believe that the learning rules applicable in both situations might be similar (Greggers and Mauelshagen, 1998; Mauelshagen and Greggers, 1993).

In invertebrates other than the honeybee, there is one study on olfactory blocking in the terrestrial mollusc Limax maximus (Sahley et al., 1981) and one on the aquatic mollusc Hermissenda crassicornica (Rogers and Matzel, 1995). Unlike the situation in honeybees, where a number of negative results for both inter-modal (Couvillon et al., 1997; Funayama et al., 1995; Gerber and Smith, 1998; for a review, see Bitterman, 1996) and intra-modal (the present study) blocking have been published, no negative results have been published on blocking in either Limax maximus or Hermissenda crassicornica. However, the studies of Sahley et al. (1981) and Rogers and Matzel (1995) were not balanced for the conditioned stimuli employed, and both studies were conducted with unpaired controls; this might have introduced the same potentially confounding variables as were demonstrated in experiment 1 of the present study and from which previous reports of olfactory blocking in honeybee classical conditioning (Smith, 1996, 1997; Smith and Cobey, 1994; Thorn and Smith, 1997) might have suffered. Further work should try to avoid these potential confounding effects to yield more convincing evidence for blocking.

In the absence of other studies, the question of the generality of blocking across sensory modalities, across training paradigms and across animal phyla remains a challenging and, to some extent, controversial issue. It is important to resolve this issue because demonstrations of blocking are critical for the implementation of cognition-like concepts of attention (Mackintosh, 1975; Smith, 1996) or of expectation and prediction (Rescorla and Wagner, 1972; Sutton and Barto, 1981) into associative learning theory.

We gratefully acknowledge help and support from the following friends, colleagues and institutions: Sylvia Lee, Beatrice Pöschel, Esther Witsch, Niels Plath and Daniel Wüstenberg for equally enthusiastic and reliable help with some of the experiments; Till Faber, Uwe Greggers, Frank Hellstern, Dirk Müller and Gregor Schulz for continuous discussions; Brian H. Smith (Columbus, OH, USA) for many discussions, comments on earlier drafts of this manuscript and the opportunity to perform experiment 1 in his laboratory. That visit was partially financed via Grant NIMH (1 R29 MH 47984). Frank Hellstern, Martin Giurfa and Jerry Rudy helped with comments on earlier drafts of this manuscript. The Studienstiftung des Deutschen Volkes and Berlin-Brandenburgische Akademie der Wissenschaften are thanked for dissertation stipends to B.G. Special thanks are due to Randolf Menzel for encouragement, discussions, comments on the manuscript and support. This paper is dedicated to the memory of Dr Juliane Mauelshagen.