The performance of cleaner wrasse, Labroides dimidiatus, in a reversal learning task varies across experimental paradigms

Spatial task and reversal learning

The experimental paradigm used in our experiments was a spatial task, whereby fish had to learn to find a food reward based on its location (left or right) in their home tank. The methods were modeled after a previous study designed to compare the performance of cleaners with closely related species (Gingins & Bshary, 2016). The basic experimental protocol was as follows (Fig. 1): subjects were simultaneously presented with two identical Plexiglas plates, placed next to each other approximately 10 cm apart. Between the two plates, a vertical Plexiglas partition was inserted to ensure fish could access only a single plate and allow the experimenter to determine when a definitive choice had been made. One of the two plates had an accessible food reward (mashed prawn) smeared on the back, whereas the second plate offered no food reward. At the beginning of each day of experiments, the tank was divided into a holding (approx. 1/3 of tank length) and an experimental (approx. 2/3 of tank length) compartment using an opaque barrier (Fig. 1), and the fish was given 30 min to acclimate. For each trial, the barrier was lifted to allow the subject full access to the tank and the plates. A choice (left or right) was noted when the tip of the fish’s snout first passed the threshold of the Plexiglas partition. The fish was given approximately 5 s to eat the food item and/or explore the experimental compartment before the barrier was placed back in the tank. The fish was kept in this holding compartment until the next trial. The trial was considered null (not taken into account) if the fish did not make a choice within 5 min. Fish were tested 20 times per day (two sessions of 10 trials) for a maximum of 100 trials (10 sessions). The task was considered solved when an individual chose the rewarding plate at least 9/10 times within a single session, 8/10 times in two consecutive sessions, or 7/10 times in three consecutive sessions. The location of the rewarding plate was randomly assigned to the right-hand side of the tank for half of the fish (n = 16), and to the left-hand side for the other half (n = 16). The location of the rewarding plate was constant throughout all the trials. All individuals that solved the initial spatial task within 10 sessions were further tested in a reversal task. Here, food plates were changed, the location of the rewarding food plate was reversed, and trials proceeded as above. The procedures for the reversal experiment were the same as described above for the initial spatial task.

Experimental treatments

Thirty-two cleaner wrasse were assigned to one of four variations of the spatial task described above (eight individuals per treatment). In the first treatment (“clear-lift”), the vertical partition placed between the two plates was transparent (“clear”). Thus, fish could see both plates at all times, and did not necessarily perceive the partition. As soon as the fish chose a plate, the second one was removed from the tank (“clear-lift”), which prevented the fish from accessing both plates during the trial (see Fig. 1). The fish were able to see the unselected plate being removed. Treatment two (“clear-stay”) used the same clear partition as in the first treatment. However, here, the experimenter only removed the second plate when the incorrect choice (i.e., the unrewarding detractor plate) was chosen. When an individual chose the correct rewarding plate, the detractor plate remained in the tank (“stay”), and could be inspected by the cleaner. In treatment three (“opaque”), the partition separating the two food plates was made of opaque Plexiglas, which prevented the fish from seeing the second plate once a choice had been made. Here, the experimenter also removed the unchosen plate as in the “clear-lift” treatment, but the fish could not see this removal happening. These three treatments were all conducted in white plastic aquaria (L = 37 cm; W = 29 cm; H = 30 cm). Treatment four (“large tank”) was carried out as described in the “clear-lift” treatment, but in a longer (L = 62 cm; W = 26 cm; H = 37 cm), glass aquarium. This setup allowed for a larger distance between the holding compartment and the plates, and thus the fish had to swim further before making a choice. Water height was maintained at approximately half of the tank height (∼15cm) in all treatments.

Statistical analysis

The number of trials to complete the task is a right-censored (i.e., maximum 100 trials), ordinal variable. Therefore, we used survival analyses to compare the number of trials needed to solve the spatial task and reversal task among the four treatments (see Gingins & Bshary, 2016). Our data did not meet the assumption of proportional hazards (assessed with cox.zph() from the “survival” R package), and thus non parametric log-rank tests were performed. Post-hoc pairwise comparisons were made for the reversal experiments (pairwise_survdiff() function from the “survminer” R package), and P-values were adjusted using the Benjamini–Hochberg method (Benjamini & Hochberg, 1995). All statistics were performed in R 3.3.0 (R Core Team, 2013). The relationship between individual’s scores in the initial and reversal experiments was tested with the Kendall rank correlation coefficient. The packages “survival” (Therneau, 2014) and “survminer” (Kassambra & Kosinski, 2018) were used for the survival analyses. All data and code for the analyses are deposited in the figshare data repository (https://doi.org/10.6084/m9.figshare.5032334.v4).

Increased distance to choice

Increasing the distance swam by the cleaners before a choice was made (i.e., large tank treatment) improved performance in the reversal learning task relative to the clear-lift and clear-stay treatments (Fig. 2B). This result parallels the findings that rats perform better in the radial arms maze with longer arms (Brown, 1990; Brown & Huggins, 1993). In this context, it appears that rats use a lax choice criterion for visiting shorter arms, thus making them more likely to revisit shorter arms (i.e., display a less accurate memory of the previously visited short arms). An increased effort to make a choice is known to improve accuracy in animals (see Kamil & Balda, 1990; Zentall, Steirn & Jackson-Smith, 1990; Bednekoff & Balda, 1997). A possible explanation is thus that visiting longer arms increased the investment before reinforcement, which could have lead rats to use a stricter choice criterion towards the long arms compared to the shorter ones (Brown & Huggins, 1993). Similarly, when the effort to reach certain arms was increased by placing barriers (rats: Roberts & Ilersich, 1989) or ropes (pigs: Laughlin & Mendl, 2004) over which animals had to climb before reaching the reward, test subjects were less likely to revisit such arms. There is a possibility that cleaners also use a lax choice criterion when distances are short. Nevertheless, in our experiments, cleaners were simultaneously presented with two plates located at an identical distance and could only “visit” one of them in each trial, hence there was no difference in investment to reach either plate, nor the possibility to revisit the same plate in the same trial. It thus seems unlikely that the choice criterion based on the investment to reach the reward played a major role here.

In our study, cleaners were conditioned to associate the removal of the barrier with the presence of a food reward. Consequently, cleaners typically darted out of the holding compartment immediately after the removal of the barrier. A longer travel distance between the holding compartment and the food plate translates into a longer computation time between the visual input of the task (i.e., plates) and the decision made. Trade-offs between speed and accuracy in both individual and collective decision making have been documented across a range of taxa (see Franks et al., 2003; Chittka, Skorupski & Raine, 2009; Latty & Beekman, 2011). Although performance was similar across all treatments in the initial spatial task, it is possible that this longer computation time provided to fish in the large tank treatment facilitated the break down of a previously learned association, which is a prerequisite for success in the reversal task. Interestingly, a transparent rather than an opaque barrier between the holding and experimental compartments could also allow for a longer computation time between the visual input and the choice of the plate, yet experiments with guppies failed to find an effect for this parameter (Gatto et al., 2017).

Opacity of the separating barrier

We found that the use of an opaque partition to separate the left from the right plate significantly improved the performance of cleaners in the reversal task (Fig. 2B). Interestingly, one could argue that having visual access to the unchosen plate via a transparent partition would facilitate learning, since the fish directly observes the removal of the rewarding plate following an incorrect choice, which should reinforce the negative association. However, transparent materials, such as the partition we used, are virtually non-existent in the natural world, and animals might have difficulties perceiving such a solid, yet transparent, object. In our experiments, we attributed a choice to the moment the fish’s head passed on one side of the partition. If the individual was unable to perceive the transparent material, this criterion may not have been appropriate for determining a true choice by the fish: an individual may have attempted to approach the rewarding plate from the opposite side (i.e., from the side of the non-rewarding plate) and consequently run into this invisible partition, a behavior we regularly observed during the experiments. This scenario is likely given that fish generally prefer to swim close to structures such as tank walls, especially in situations which may induce anxiety (Maximino et al., 2010). Thus, an individual might have “known” which plate offered the food reward, but effectively “chose” the detractor plate because it approached the plates from the wrong side of the aquarium, not realizing the transparent partition would prevent it from accessing its preferred choice. This issue is intuitively more likely to occur in the reversal experiment. During the spatial task, individuals would have become conditioned to approaching the plates from the side initially offering the reward. Although the rewarding side was switched in the reversal experiment, the side preference for the approach would likely carry over from the first experiment even if fish did not receive a reward. This could explain the improved performance of cleaners in the opaque partition treatment, where the fish were able to clearly perceive the separation between the options. In other words, our criteria for what constituted a correct choice may have favored opaque barriers. If the choice criteria had been for the fish to touch the plate, it is possible that learning speed would have been similar for both opaque and transparent barriers, since many subjects did not actually touch the non-rewarding plate but went straight against the partition in the transparent treatments.

The role of ecology

In nature, client fishes that have access to several cleaning stations are less likely to return to the same cleaner if they had been ignored during their previous visit (Bshary & Schäffer, 2002). Ignoring a client can thus have negative consequences for cleaners, and we had hypothesized that cleaners might also associate the removal of the unchosen plate as the loss of a foraging opportunity (i.e., a potential client leaving the cleaning station to seek service elsewhere). Removing the unchosen plate in view of the cleaner might have thus impaired their ability to associate the positive feedback of the reward with their decision. As a result, we expected that removing the unchosen plate only when individuals made the wrong decision (i.e., the clear-stay treatment) would facilitate learning for cleaners in this task. We did not find evidence supporting this prediction, suggesting that the nature of cleaner-client interactions had little influence on performance in the spatial task. In nature, cleaners do not face noteworthy spatial challenges other than during navigation within their limited home range, and, thus do not perform better in spatial tasks than do other wrasse species (Gingins & Bshary, 2016). It would therefore be interesting to repeat our manipulation in an ecologically relevant task: cleaners appear to form negative associations between their decisions and the departure of a client fish in nature (Bshary & Grutter, 2002b), or feeding plates in the lab (Bshary & Grutter, 2005). One could thus repeat the study with different colors or patterns to identify the rewarding plate, rather than by its location.

Additional sources of variation

Several additional factors might contribute to explain the results we found in this study. First, we use a relatively small sample size (eight individuals per treatment). Age, sex, rearing conditions and previous experience can affect individual performance in cognitive tasks (Thornton & Lukas, 2012; Lucon-Xiccato & Bisazza, 2017). Such inter-individual differences might impact the outcome of experiments when a limited number of subjects are tested, since small sample sizes are less likely to capture the whole spectrum of variation in performance (Thornton & Lukas, 2012). In cleaners, variation in performance in certain cognitive tasks has been documented across individuals from diverse developmental environments (see Wismer et al., 2014). Here, we controlled for this potential source of variation by testing only cleaners caught in similar habitats, but we cannot rule out that other ontogenetic factors or effects due to age might have an effect on our results. Nevertheless, we believe that the large differences we observed in the reversal experiment cannot be attributed exclusively to the small sample size used. Understanding the ecological factors driving inter-individual variation in physiological, behavioral and cognitive traits observed in this system is an interesting question in its own right which merits further investigation.

Second, while we had controlled for odour cues in a previous study using the same setup (Gingins & Bshary, 2016), we did not control for odour cues in the present study. Even though the odour of the food reward could potentially affect performance in a spatial task, we remain confident that odours did not influence our results for several reasons: (1) all treatments used in our comparison were performed in the same manner. An odour cue from the correct plate should help a fish solve the task, and would thus bias the results towards the null hypothesis (i.e., finding no effect of treatment). (2) If cleaners used odour cues, this spatial task should have been straightforward for them to solve, yet in two treatments (i.e., clear-stay and clear-lift) all fish failed to solve the reversal task. The fact that we observed low performance in this task despite the potential help of odour cues suggests that odour did not have an important influence on a fish’s performance. (3) We can compare the results of this experiment with that of a previous one in which odour cues had been controlled. The performance of L. dimidiatus in Gingins & Bshary (2016), which correspond to our clear-lift treatment and where odour cues were controlled for, is similar to that of our fish. Gingins & Bshary (2016) had four out of eight individuals solve the task within 60 trials. Using the same criteria (60 trials), we observed five out of eight fish solve the task in our study. Therefore, it seems unlikely that odour cues helped the fish solve this spatial task.