Stimulus stringing by pigeons: Conditional strings

Animal Learning & Behavior1983,11 (1), 19-26

Stimulus stringing by pigeons:Conditional strings

W. KIRK RICHARDSON and JEFFREY A. KRESCHGeorgia State University, Atlanta, Georgia

Pigeons were trained to produce one serial list in the presence of a green background cueand another serial list in the presence of a red background cue when the items for both seriallists were presented on each trial. This demonstrated a combination of serial learning andconditional discrimination learning not previously shown in pigeons. Specifically, when pre-sented with four geometric forms, ABC D, in random locations of a five-key display, thepigeons learned to peck ABC when the background was green and A B D when the back-ground was red. Accuracy on the conditional string ranged from 73% to 85%. Transfer testsusing different locations of the stimuli on the keys showed positive transfer, thus ruling outlearning of specific locations as the basis of the accurate performance. Above-chance perfor-mance was maintained when the conditional colors were presented only on the key that didnot contain one of the serial stimuli. The results are interpreted in terms of a chaining modelthat postulates that the sequential selections were controlled by cues produced by both onsetof the trial and prior selections within the trial.

Serial learning has long been a research topic,especially in the literature on humans. Currently,there are a number of paradigms being used to studyserial learning in lower animals. Sands and Wright(1980) have demonstrated that rhesus monkeys havea remarkable capacity to learn a long serial list.Capaldi's theory of partial reinforcement has beenextended to the learning of monotonic and non-monotonic series of quantities of food rewards(Capaldi, Verry, & Davidson, 1980). Hulse (1978)has also carried out extensive studies of serial patternlearning for series of food quantities. Studies ofspatial memory in the radial arm maze (Olton, 1978)can also be viewed as using a serial task.

The pervasiveness of serial behavior means thateach organism generally engages in different behav-ior sequences at different times. The particular se-quence active at a given time would presumably bedue to some stimulus, either discriminative or elicit-ing, that controls the behavioral sequence. However,the paradigms used to study serial behavior generallyinvolve only one such sequence for a given subject.When different sequences are studied, they are usuallygiven to different groups of animals.

The present study extends a paradigm used in ourlaboratory to study single sequences in pigeons(Richardson &: Bittner, 1982: Richardson &: Warzak,

This study was based on a thesis submitted to the Departmentof Psychology, Georgia State University, in partial fulfillment ofthe requirements for the MA degree by the second author, Wewish to thank the staff of the Georgia State University ComputerCenter for technical assistance. Reprints may be obtained fromW. Kirk Richardson, Psychology Department, University Plaza,Atlanta, Georgia 30303.

19

1981) to the study of performance by individual or-ganisms on two different sequences. In our earlierwork, the serial items were colors projected on re-sponse keys. Each color was assigned to a serial posi-tion in the stimulus string (sequence). If we called thefirst color A, the second color B, etc., then the task re-quired the subject to peck ABC D in that order inthe four-stimulus string condition. The location of thecolors on the keys was varied randomly from one cor-rect trial to the next, and a between-trials correctionprocedure was used to insure that the subject learnedeach array. Correctly selected colors changed fromdim to bright for the remainder of the trial. Thus, acorrectly selected stimulus was explicitly "tagged."The trial terminated after a correct sequence or oneincorrect selection.

The present study extended the earlier work in twoways. First, the items consisted of lexigrams (geo-metric drawings) projected as white illuminated lines.All earlier work had used colors for the serial items.Because color was known to be a very salient cue forpigeons, we wished to see if pigeons could learn astring when the stimuli seemed to have no specialsalience. The second, and more important, extensionwas the requirement that the pigeon learn two strings.The elements for both strings were provided, and anadditional cue indicated whether string 1 or string 2was correct on a given trial. The additional cue wasone of two colors, red or green, projected as thebackground color for the lexigrams. The task wasthus a conditional discrimination, in which the condi-tional behavior was the selection of a stimulus string.Because the correct sequence of lexigrams was indi-cated by the background color and because there wasno physical similarity between the background cues

Copyright 1983 Psychonomic Society, Inc.

20 RICHARDSON AND KRESCH

EXPERIMENT 1

Figure 1. Scbematic drawings of the lexignms used as serialitems in tbe strinp.

and the lexigrams, the task can be referred to as aconditional, symbolic, stimulus-string task.

Experiment 1 of the present study answered twoprimary questions. First, could pigeons learn a stim-ulus string when the elements of the string did not

h~ve any special salience for pigeons? Second, couldpigeons learn to select one of two possible stimuliin the variable serial position of a string when thecorrect stimulus selection was cued by the value ofan environmental cue? If the pigeons learned thistask, then more molecular analyses would help tounderstand the learning processes involved.

An Interdata 732 computer controlled the experimental task anddata collection.

Proc~ure.After adaptation and magazine training, the subjectswere trained to peck Stimulus A on a black background using themethod of shaping by successive approximations. Pecks to Stim-ulus A (the first stimulus in the sequence) produced food and thentrial termination. In all phases of the experiment, any incorrectresponse produced a .S-sec tone and then trial termination, whereas

corre~t stn~gs produced food..When the subjects were reliablytracking Stimulus A, the two-stimulus sequence (A B) was intro-duced. When the subjects were reliably pecking the sequence A Bthe three-stimulus sequence was introduced. '

.In t~e three-stimulus condition, there were two groups of threestimuli presented to each subject: ABC and A B D. When ABCwas presented, the background was green. When A B D was pre-sented, the background was red. Thus, the color cue was corre-lated with the stimulus set presented, but color was a redundantcue, since both C and green were present when C was correctin Serial Position 3 and both D and red were present when Dwas correct in Serial Position 3.

A discrete-trials procedure with a between-trials correction con-tingency and a 2-sec intertrial interval was used in all conditions.At the beginning of a trial, the array was presented at the dim in-tensity on the bottom row of keys. The top row of keys was notused. Each correct selection of a stimulus caused the stimulus tochange from the dim to the bright intensity for the duration of thetrial.

The different possible permutations of the stimuli on the fivekeys were called arrays. For the one-stimulus condition, there were5 possible arrays, since the stimulus could be placed on each ofthe five keys. All 5 arrays were used in training. For strings withtwo or more stimuli, 10 arrays were chosen from the possible per-mutations. The 10 arrays were chosen so that each stimulusoccurred on each key equally often, that is, so that stimuli andkeys were not confounded.

After a correct trial, the array for the next trial was chosenfrom a randomized-blocks order of the 10 arrays, with 20 arraysper block. After an incorrect trial, the same array was presentedagain (a correction procedure). However, the subject had to startat the beginning of the string (Stimulus A) regardless of the loca-tion of the error on the prior trial. Thus, the subject had tocomplete correctly an array before being presented with the nextarray. This correction procedure prevented the use of an error to"skip over" a difficult array. Each session lasted for 140 correcttrials or 45 min (60 min during the initial training), whichevercame first.

Because the subjects made many errors during the initial ses-sions of the three-stimulus sequence, two procedures were usedto facilitate acquisition. First, some shaping was employed bymanually delivering food for components of the string. For exam-ple, pecks to A would be shaped for a subject that tended to peckC in the first serial posi~ion. Second, for four of the subjects,arrays that were more difficult were removed until the subjecthad mastered the easier arrays. When accuracy (number of cor-rect strings/total number of strings) showed no systematic changeover five sessions, the conditional string task was introduced.

In the conditional string task, the arrays contained all fourstimuli (A, B, C, and D) and one dark key. Within each block of20 correct trials, each of the 10 arrays was presented twice.Within a block, the background was green once and red once foreach array, and the within-block order was determined randomly.The background color served as the conditional cue. When it wasgreen, the correct order was A, B C. When it was red, the correctorder was A B D. This phase was terminated after 34 or 36 ses-sions.

Finally, the subjects were given two transfer tests. In each trans-fer test, 10 new arrays were used in place of the original arraysfor 10 sessions. All other details were identical to the previouscondition.

D

D

*'

c

c

nB

+

B

=a:::

A. Stimulus Set A.

B. Stimulus Set B.

A

Method~ubj~ts. The subjects were eight naive racing homing pigeons

maintained at 7511fo of their free-feeding weights. Three subjectswere assigned randomly to Group A (Stimulus Set A) and fivesubjects were assigned randomly to Group B (Stimulus S~t B).

Apparatus. Two test chambers had inside dimensions of SO x36 x 36 em (L x W x H) and a jeweled light centered on the rearwall 4 cm from the ceiling. Opposite to the rear wall was analuminum-alloy false wall with two rows of five 2.5-cm-squareopenings. The top row had its upper edge 7 em below the topof the chamber and its left edge 14 cm from the left wall. Thebottom row was 4.5 em below the upper row. There was 1.5 cmbetween the sides of adjacent openings. A Scientific Prototype foodcup was located on the lower left of the front wall, 1 em abovethe floor and 2 em from the wall. The food cup could be illu-minated by a hooded lamp centered .5 ern above the food cup. A75-dB (re 20 SPL) white noise and a ventilation fan providedauditory masking.

An Industrial ~lectrOnicsEngineers Series 10 in-line display cellwas located behind each opening. Each display cell containedthree colors and nine geometric forms, the same elements used toconstruct the lexigrams of the Yerkish language (von Glaserfeld,1977). Two sets of four stimulus cues (lexigrams) were formedfrom randomly selected pairs of the nine geometric forms (Fig-ure 1). The lexigrarns could be projected against background

col~rs of black (i.e., no color), green, or red, and the color-lexigram compound could be illuminated at two intensities dim(4 V) and bright (6.5 V), by No. 44 miniature lamps. Trans~arentLexan paddles located between the openings and the display cellscould be operated by a static force of 5-15 g (.005-.05 N) throughan excursion of .5-1.0 mm.

STIMULUS STRINGING 21

Fiaure 2. Tbe probability for eaeb type of error for tbe lastsessions of tbe tbree-stimulus striag, tbe first session of the con-ditional string, and tbe last sessions of tbe conditional string.

are forward errors. Selections backward in the se-quence (e.g., the sequence A B A) are backwarderrors. Selections of the same stimulus twice in suc-cession (e.g., A B B) are repeat errors. Any selectionof a dark key (a key not containing a stimulus)during a trial is a dark-key error. The fifth errortype, the intrusion error, can occur only in a condi-tional string. This error is the mismatch of the vari-able stimulus and the background color, that is, Dwhen the background is green and C when the back-ground is red. Since the number of opportunities tomake different types of errors varies with error type,string length, serial position, etc., the probability ofeach error type was computed for each subject foreach session.

For each session's data, the probability of an errortype was computed by dividing the number of occur-rences of the error type by the number of oppor-tunities for that error type to occur (see Richardson"Bittner, 1982, and Richardson" Warzak, 1981, formore details). Figure 2 presents the mean error typeprobabilities for the last three sessions of the three-stimulus problem, the first session of the conditionalstring, and the last three sessions of the conditionalstring. The error probabilities of the four error typespresent in both the three-stimulus problem and theconditional string showed little change from the lastof the three-stimulus string to the first session of theconditional string. An analysis of variance of for-ward, backward, repeat, and dark-key errors for theend of the three-stimulus string and the first day ofthe conditional string showed no effect of condition[F(l,7) < 1), an effect of type of error [F(3,21)=118.29), and no interaction of condition and errortype [F(3,21) = 1.64). The drop in accuracy at the be-ginning of the conditional string was due to the highprobability of intrusion errors. By the end of the con-

I - Last 3 sessions of three-stimulus stringII - First session of conditional string

III . Last 3 sessions of conditional string

Dark-Key Intrusion

ERROR TYPE

I II III

Forward Backward Repeat

,012

.010

>-...

::i .008iiicta:l .0060II:Q.

.004

.002

Results and DiscussionAccuracy is defined as the number of correct

strings divided by the sum of the number of correctand the number of incorrect strings. The chance levelof accuracy of a three-stimulus string is the probabil-ity of randomly picking the correct key out of fiveavailable keys on three successive selections, which is.008. A more conservative chance model would havethe subjects randomly select from the dim keys onlyunder the assumption that they had learned thebright-dim discrimination. This model would give achance accuracy of .167 (l/3xll2xl/l) for thethree-stimulus string and .042 (1/4 x 1/3 x 1/2) forthe conditional string. All of the accuracy levels pre-sented below were clearly' above chance.

Accuracy was 64.9070 (range 52.4% to 75.3%) onthe last session of the three-stimulus string anddropped to 38.4% (range 27.7% to 75.3%) on thefirst session of the conditional string. The drop wasstatistically significant [F(l,7) = 85.83). All statisticaltests used the .05 level of significance. Training onthe conditional string raised accuracy to 78.4070(range 72.5% to 84.8%) on the session before thefirst transfer test.

The best measure of transfer uses only the firstpresentation of each new array on the two back-ground colors, because responding to subsequentpresentations may be influenced by the consequencesof the responses to the prior presentations. Theappropriate comparison for the first presentationaccuracy on the new arrays is the accuracy for thefirst presentation on the old arrays within the pre-vious session. This controls for possible warm-upeffects.

The first exposure data for the first transfer showeda significant drop, 69.4% to 51.3% [F(l,7) = 11.52),although the accuracy remained above chance. Thesecond transfer test did not show a change in accu-racy [F(I,7)=2.71).

Introduction of the conditional string task pro-duced a disruption of behavior followed by quickrecovery. One possible explanation of the effect isthat the subjects did not learn about color during thethree-stimulus problem, that is, that they failed toassociate. Colors were a redundant cue in the three-stimulus problem, since the subject could learn topeck ABC when presented with those three stimuliand to peck A B D when presented with that stimulusset. Only when the conditional string was introduceddid color become a relevant cue. The failure-of-association hypothesis would predict that the addi-tional errors that occurred upon the introduction ofthe conditional string would be the incorrect choicesof one of the variable stimuli. In order to test thishypothesis, an analysis of error types was performed.

In this task, there are five types of errors. Selec-tions forward in the sequence (e.g., selecting C first)


Figure 3. Probability of an intrusion error in each serial positionas a function of blocks of 20 correct trials.

.140... SERIAL POSITION\

.120 \ 0 ___ Firsta:: \\ ---- Second0 \

a:: \ .. ---_ Third> a:: \

.100 ,t- W . \::J z

,\I \iii 0 " , en .080

"\'a:I ~0 a:: tt "a:: t- " \.Q, :!: .060 I ....u. . ....0.040 ' ........." ..._~~~ ...

....

.020 .~. ....' ......,.c:., --o~ ..

.............. ....- .....~--__

the conditional string task used in Experiment I. Next, the back-ground ,cue was removed from lexigrams C and D of each array,so th~ fixed elements of the string, but not the variable elements,contained the conditional cue. After six sessions, the conditionalcue was placed, once again, on all lexigrams. After 6 more ses-sions, the conditional cue was again removed from the variableelements for 4 additional sessions.

In the next condition, the background cue was removed fromall lexigrams and was presented on the key that did not containa lexigram. After 12 sessions under this condition, the background

c~e was removed from the no-lexigram key so that the displaydid not contain a conditional cue in any location (an empiricalmeasure of chance accuracy for the conditional discriminationpart of the task). After 4 sessions, the background cue was re-turned to the no-lexigram key for 10 sessions.

Results and DiscussionA~c~racy for Sessions 16-62 is plotted in Figure 4.

Statistical analyses of the changes in conditions wereperformed by comparing the last session of each con-~tion ~ith the first session of the following condi-~10n (USl~~ the .0' level of significance). All changes10 conditions produced a statistically significantchange in accuracy.

When the colors were first removed from thevariable-lexigram keys, there was a large drop in ac-curacy [F(l,7) = 82.13], followed by recovery overfour sessions. Replacing the colors on the variable-lexigram keys also produced a small drop in accuracy[F(I,7) = 8.43], whereas removing the colors from thevariable-lexigram keys for the second time was fol-lowed by a small increase in accuracy [F(l,7)=7.15].

These data clearly show that the background colorwas not an integral or a necessary part of thevariable-lexigram stimuli for the subjects; otherwise,they could not have performed well in the no-color-on-the-variable-lexigram-key condition. Evidently,the subjects saw the stimuli as lexigrams on a coloredbackground, as did the experimenters.

The effect of removing the colors from the lexi-gram keys and placing them on the no-lexigram keywas large, immediate, and lasting [F(I,7) =439.84].There was some recovery during the first exposure to

location of conditional cue

ARC.!> A.B I A.B.{'.D A.H no 1""''1r..m I"'l no 1.... ;9rdlm k"l

loin ,e.: ,e_.: e., :-._"./_1 I::"~:': :,;...... ! / I ,.,/~ '" iii (.. ',i:,1c:x:: "II : : I:'o :. IU ! l I !~ '.11 :: I :._,_,-I.lJ :: I .~ .-.-./ ,-,,,; [e.."-U ): : _,.-- : Iffi to :: I'::I l/~"':Q. i [ : :

!) .- !!: i ! !!

[" I '" ,I .' _ _ _ _ ,t" t, t ,I", ,I. , , , , , , , ! ,If> 1M .18 40 42 -t4 4f> 414 ,'iO 52 54 56 58 flO &2

SESSION

Figure 4. Accuracy during tbe conditions of Experiment 2.


this condition, but accuracy never reached 500/0. Thefollowing condition had no colors present to serveas conditional cues. It was a conditional discrimina-tion without conditional stimuli, an empirical mea-sure of chance for the preceding condition. Note thatthis is a measure of chance for the conditional aspectof the task, not a total random-responding chance.The subjects still had the lexigrams to serve as dis-criminative stimuli for responses in Serial Positions1 and 2. The absence of the conditional cue makesSerial Position 3 a coin toss (assuming selection of Cor D). The expected value for accuracy would be .50only if selections in Serial Positions 1 and 2 werealways correct, which they were not. Comparison ofthe chance condition with the preceding [F(1,7) =17.70] and following [F(1,7) = 33.07] colors-on-the-no-lexigram-key condition showed that the subjectswere performing above chance when the colors wereon the no-lexigram key. The last colors-on-the-no-lexigram-key condition had an increase in accuracybetween the first and last sessions [F(l,7) = 24.08].Improvement would probably have continued iftraining had been continued. Using a single eight-stimulus string, we have found a slow increase inaccuracy across 80 training sessions (Mitchell Cole& Richardson, Note 1). ' ,

The low level of performance observed when thecolors were not on the lexigrams was probably dueto attentional factors. In matching-to-sample tasksaccuracy is typically very low unless the subject i~required to peck, and thus observe, the sample priorto making the matching response (EckermanLanson, & Cumming, 1968). The high accuracy o~the two conditions in which the colors were on someor all the lexigrams was probably due to the factthat the subjects had pecked, and thus observed theconditional cue prior to choosing from the variablelexigrams. If a peck to the no-lexigram key had beenrequired when it alone contained the colors then

. . 'an increase 10 accuracy might have been observed.Other ways of forcing attention to the conditionalcues might also be effective. For example, the lexi-grams could be lighted without a conditional cue,and any response could be counted as an error. Aftera variable time without a response, the conditionalcue could be presented away from the lexigrams.The presence of the conditional cue would signalboth the beginning of the trial and which variablelexigram would be correct. If the subjects havelearned to attend to the conditional cue to avoidpremature responses, then the cue should also con-trol choice of the variable lexigram.

GENERAL DISCUSSION

The results of these studies show that the pigeonssolved the two-string problem as a conditional dis-crimination, that is, that the selection of three offour stimuli came to be controlled by background


color as the result of differential reinforcement ofthe strings in the presence of the two backgroundcolors. There are thus two types of discriminativebehavior to be explained: the selection of the cor-rect string and the selection of the correct sequenceof items within a string. The selection of the correctstring reduces, in the present experiments, to the selec-tion of the correct item in Serial Position 3, theonly variable position here. The selection of thecorrect item in Serial Position 3 could be due tosimple paired-associate learning. The pigeon pecksA, then B, and then C or D, depending on thestate of the conditional stimulus (if green, peck C;if red, peck D). When the conditional color wason the lexigram keys, the pigeons had just peckedand thus observed the color while pecking A and B,which facilitated the subsequent selection of the cor-rect lexigram in Serial Position 3. When the colorwas off the lexigram keys, the pigeons were notrequired to peck the conditional color prior to se-lecting C or D and thus may not have observedthe color, which resulted in a lower level of accuracy.

Different explanations have been offered for thelearning of the serial sequence in this type of para-digm. Straub, Seidenberg, Bever, and Terrace(1979) took the position that the pigeon learns arepresentation that allows it to keep its place whileworking through the required series of items.Richardson and Warzak (1981) suggested a chain-ing model, in which there are stimuli present ateach point in the sequence that may act as dis-criminative stimuli for the next selection. For thepresent study, the chaining model would work asfollows. Trial onset would produce discriminativestimuli controlling a high probability of the selec-tion of A; the selection of A would produce stim-uli controlling a high probability of the selectionof B; and the selection of B would produce stimulithat, in combination with the conditional color,would result in a high probability of the selectionof C or D, whichever was appropriate. The stimuliproduced by a selection that serve as discriminativestimuli for the next selection could be external, in-ternal, or both. In the present study, a likely can-didate for part of the discriminative stimulus com-plex would be the bright state of the stimulus thathad just been selected. Recall that the correct se-lection of a stimulus resulted in the change of thatstimulus from the dim to the bright state. Althoughthis feedback would be expected to be an impor-tant part of the discriminative stimulus, it is notnecessary (Richardson & Bittner, 1982; Straub &Terrace, 1981), since distinctive natural consequencesof pecking a specific stimulus would also be present.As would be expected if a salient part of the dis-criminative stimulus was removed, the Richardsonand Bittner study showed lower accuracy for the

condition in which the experimenter provided nofeedback.

Straub and Terrace (1981) have demonstratedthat pigeons trained to select a four-stimulus string,ABC D, can perform above chance on test trialsconsisting of an abbreviated form of the string, forexample, A C or A D. Our chaining model can ex-plain these data also.

In order to decide which responses will be pre-dicted by the chaining model, we must consider howthe discriminative stimuli come to control the nextresponse, a question whose answer must considerthe training procedure whereby the pigeons came tolearn the task in the first place. The procedure isthat the selection of A in the presence of the dis-criminative stimuli existing after trial onset is fol-lowed by conditioned reinforcement and, after somewithin-chain delay, primary reinforcement when acorrect sequence has been emitted. Our assumptionis that the cues present after trial onset becomestrongly associated with the selection of A. In addi-tion, these same cues will also be associated, to alesser degree, with the selection of B, C, and D.These associations will occur to the degree that trialonset cues are also present when the later stimuliare correctly selected. There are two reasons thattrial onset cues could be present when the later itemshave been correctly selected. First, since the suc-cessive selections occur over a short time period,the trial onset cues may not have completely fadedfrom memory when the later selections occur. Second,the cues produced by the selection of A, B, C, andD may all have some elements in common withthe cues produced by trial onset, a common-elementsmodel of stimulus generalization. In general. thereshould be a gradient of association between the cuespresent at any point in the chain and all selections.Thus, after training with ABC D, the subject wouldselect A first on an A D trial because trial onsetcues are strongly associated with the selection of A.The subject would then be likely to select D nextbecause the cues present after the selection of Aare associated with the selection of D as well aswith the selection of Band C, which were notpresent on A D trials.

Straub and Terrace (1981) found that the pigeonsperformed well on the abbreviated strings and thatin the case in which the abbreviated string con-sisted of A and one other item, the accuracy in-creased with the number of items from the train-ing string that were omitted from the two-stimulustest string [e.g., the A B test string had no itemsomitted, whereas the A D test string had two items(B and C) omitted]. This increase in accuracy seemscounterintuitive and unpredictable from a chaininganalysis. Why should accuracy be lower on A Btest trials, in which the two items were in the same

relative positions as in training, than on A D testtrials, which consisted of two items the pigeons hadnot been trained to peck successively? In order tounderstand how this distance effect is in fact pre-dicted from our chaining model requires closer exam-ination of the procedure used by Straub and Terrace(1981). That procedure did not have experimenter-provided feedback and did not count repeat pecksto a correctly selected item as errors; thus, whenpresented with the pair A B, the only possible wayto make an overt error was to select B in the firstserial position. Failure to complete a trial in 15 secalso terminated the trial and was in one sense anerror. This time limit was seldom reached, and,since it was not relevant to the present analysis,it was ignored. If A was selected first, then thetrial could not end in an error, since subsequentselections of A were not counted as errors and se-lection of B, the only other stimulus present, wouldresult in a correct trial. Likewise, when the pairA D had been presented, the only possible errorwas the selection of D in the first serial position.

Now consider what happened on two-item testtrials. First, since trial onset consisted of present-ing only two of the four cues used during train-ing, the abbreviated string trial onset cues coulddiffer slightly from the regular training trial onsetcues. However, there is no reason to expect thisdifference in trial onset cues to produce a bias forany test item. When Items A and B are presentedas an abbreviated string, both should be stronglyassociated with the trial onset cues, but Item Ashould be more strongly associated. At the otherextreme, with the presentation of Items A and Das an abbreviated string, Item D should be lessstrongly associated with the trial onset cues thanItem B on the A B trials, due to the gradient ofassociation suggested above. Thus, a strongly as-sociated item (B) would compete more effectivelywith Item A on A B trials, whereas a weakly as-sociated item (D) would compete less effectivelywith Item A on A 0 trials. The weaker competi-tion on A D trials would result in a higher prob-ability of selecting A in Serial Position 1, and, aswe have shown, once A has been selected, the onlyother selection that would count is the selection ofD, resulting in a correct string. Repeat selectionsof A are not counted as errors. The higher prob-ability of selecting A in the first serial position onA D trials results in a higher accuracy for A D trials.The same reasoning would predict the accuracy onA C trials to be between the accuracy on A B andA D trials. Thus, these results of the abbreviatedstring tests are predicted by the chaining model.If repeat selections had been counted as errors inthe Straub and Terrace (1981) study, the resultsmight have been different, since the pigeons may


have been more likely to repeatedly peck A on A Dtrials,on which D weakly competes with A, thanon A B trials, on which B strongly competes withA.

Finally, Thompson and Church (1980) presentedan analysis of the language-like behavior of LanaChimpanzee, which may be relevant to the presentparadigm. Lana Chimpanzee (Rumbaugh, 1977)learned to select a string of symbols depending onthe conditional cues present in the environment. Forexample, if the three conditional cues (1) objectis present, (2) object is food, and (3) object is in ma-chine were present in the environment, then Lanacould select the stock sentence "Please machine give(incentive)." If Lana inserted the lexigram for thefood that was visible in the machine into the placeof (incentive), then she would receive the food inthe machine. We feel that the paradigm used withLana is similar to the conditional stringing paradigm.In fact, the work with Lana provided the impetusfor the present work. The Thompson and Church(1980) analysis ascribed Lana's selection of a stocksentence (stimulus string) to conditional discrimina-tion learning, while paired associative learning wasresponsible for the selection of the appropriate lexi-gram to represent an incentive in the stock sen-tence. This analysis would work equally well forthe behavior observed in the present study.

REFERENCE NOTE

I. Mitchell, Z. P., Cole, M. A., &: Richardson, W. K. Theorigin of forward errors in serial lists by pigeons. Paper pre-sented at the meeting of the Southeastern Psychological Associa-tion, Atlanta, March 1981.

REFERENCES

CAPALDI, E. J., VERRY, D. R., &: DAVIDSON, T. L. Memory,serial anticipation pattern learning, and transfer in rats. AnimalLearningcl:Behavior, 1980, I, 575-58S.

ECKERMAN, D. A., LANSON, R. N., &: CUMMING, W. W. Acqui-sition and maintenance of matching without a required observ-ing response. Journal of the Experimental Analysis ofBehavior,1968, 11,435-441.

HULSE, S. H. Cognitive structure and serial pattern learning byanimals. In S. H. Hulse, H. Fowler, &: W. K. Honig (Eds.),Cognitiveprocessa In animal behavior. Hillsdale, N.J: Erlbaum,1978.

OLTON, D. S. Characteristics of spatial memory. In S. H. Hulse,H. Fowler, &:W. K. Honig (Eds.), Cognltlveprocuses in animalbehavior. Hillsdale, N.J: Erlbaum, 1978.

RICHARDSON, W. K., &: BITTNER, B. J. Stimulus stringing bypigeons: Effect of feedback for correct selections. AnimalLearning cI: Behavior, 1982, 10, 35-38.

RICHARDSON, W. K., &: WARZAK, W. J. Stimulus stringing bypigeons. Journal of the experimental Analysis of Behavior,1981,36,267-276.

RUMBAUGH, D. (Ed.). Language learning by a chimpanzee: TheLana project. New York: Academic Press, 1977.

SANDS, S. F., &: WRIGHT. A. A. Serial probe recognition per-formance by a rhesus monkey and a human with 10- and 2O-item


lists. Journal of Experimental Psychology: Animal BehaviorProcesses, 1980,6, 386-396.

STRAUB, R. 0., SEIDENBERG, M. S BEVER, T. G., & TERRACE,H. S. Serial learning in the pigeon. Journal of the ExperimentalAnalysis ofBehavior. 1979.31, 137-148.

STRAUB, R. 0., & TERRACE, H. S. Generalization of serial learn-ing in the pigeon. Animal Learning & Behavior, 1981,9, 4S4-468.

THOMPSON, R. T., & CHURCH, R. M. An explanation of thelanguage of a chimpanzee. Science, 1980,108,313-314.

VON GLASERSFELD. E. Linguistic communication: Theory anddefinition. In D. M. Rumbaugh (Ed.), Language learning by achimpanzee: The Lana project. New York: Academic Press,1977.

(Manuscript received April 26, 1982;revision accepted for publication August 29, 1982.)

Documents

Stimulus stringing by pigeons: Conditional strings