A behavior field approach to instrumental learning in the rat: II. Training parameters and a stage model of extinction

Three runway experiments tested a stage model of extinction which postulated an orderly succession of three qualitatively different stages: habit, trial and error, and resolution. The model predicted that Stage 1 should be characterized by perseveration of habitual routes (i.e., response persistence) and the absence of competing responses; Stage 2, by an increase in investigatory behavior (response variation and hole exploration) and biting behavior; Stage 3, by a decrease in the competing responses of Stage 2 and continued increase in goal avoidance and substitution behavior (e.g., sand-digging). These predictions were largely confirmed. Further, Experiments 1 and 2 showed that, as expected by the model, continuous reinforcement (CRF) resulted in more practice of habitual routes in acquisition and greater response persistence, while partial reinforcement (PRF) resulted in more route variation and hole exploration in acquisition and greater goal persistence which was attributable to prior reinforcement of a trial-and-error coping strategy. Results of Experiment 3, which combined training trials and reward magnitudes orthogonally, supported the prediction that response persistence was positively related to training trials, and goal persistence negatively related to reward magnitudes. All three experiments demonstrated an inverted-U function in investigatory and biting behavior, as predicted by the stage model.     

Extinction of operant behavior: An analysis based on foraging considerations

In two experiments, the frequency of food reinforcement provided by variable interval (VI) schedules prior to extinction was varied. In the first experiment, two-component multiple schedules resulted in a greater number of responses in extinction in the presence of the stimulus previously associated with the richer of the two component schedules than that previously associated with the leaner schedule. In the second experiment, different groups of animals were trained on different VI schedules. Responding in extinction was analyzed into bouts of responding showing that the number of response bouts increased and the number of responses per bout decreased with decreasing frequency of reinforcement during training. These data are compatible with an analysis of operant behavior based on an analogy to processes that presumably-occur-in naturalistic foraging situations. According to this analogy, behavior associated with search for a food source (i.e., number of response bouts) and that of procurement of food from a source (i.e., responses per bout) represent aspects of behavior that are differentially strengthened by different VI schedules. Extinction serves to reveal this differential strengthening.     

Behavioral variability as a function of response topography and reinforcement contingency

Long-Evans rats were reinforced for generating variable sequences of responses on two operanda. The current sequence of four left and right responses was required to differ from each of the previous five sequences. Variability under thisvary schedule was compared with that under ayoke control schedule where reinforcement was independent of the sequences. Three different response topographies were compared: two levers were pressed in one case, two keys pushed in another, and two wires pulled in a third. Both reinforcement contingency (vary vs. yoke) and response topography (leverpress, key push, and wire pull) significantly influenced sequence variability. As is the case for operant dimensions, behavioral variability is jointly controlled by reinforcement contingency and response topography.     

The source of keypecking in autoshaping

Pigeons were trained to step on a treadle to operate a grain hopper, under the control of an auditory stimulus. Subsequently, autoshaping consisted of pairing illumination of a response key with some of the subsequent tone presentations, reinforcement occurring only after a treadle response. One control group did not have a treadle or a treadle-response requirement. A second control group received random presentations of an illuminated key. Neither paired nor random key presentations had any effect on treadlepressing. However, the paired key groups showed acquisition of pecking to the key, while the random group did not. The data suggest that pecking in the autoshaping procedure is not profitably analyzed as part of an operant sequence and it indicates that autoshaping involves a structuring of response repertoires that may well be independent of the effects of contingent reinforcement and shaping. Autoshaping seems to be governed by empirical temporal laws akin to those of classical conditioning.

     

The overshadowing of instrumental conditioning by a stimulus that predicts reinforcement better than the response

In an experiment designed to demonstrate the overshadowing of appetitive instrumental conditioning, three groups of rats were given 10 sessions of RI (random interval) training in which reinforcement was delivered 450 msec following the response. The correlated group experienced a stimulus during the response-reinforcer delay interval, while an uncorrelated control group experienced a similar brief stimulus occasionally following responses, but these were not responses that typically produced reinforcement. A no-stimulus control group did not experience the brief response-produced stimulus. The experiment was run in two replications. The first employed a light as the critical stimulus, the second a tone. Over the 10 RI sessions, subjects in the two control conditions increased their rate of responding significantly faster than subjects in the correlated condition in both replications. This finding was interpreted as an instance of the overshadowing of the acquisition of signal value by a response because of the presence of the stimulus, which, in the correlated condition, was a more reliable predictor of reinforcement than the response. A subsequent conditioned reinforcement test confirmed that, in the correlated condition, the stimulus had, indeed, become a signal for reinforcement as a function of RI training.     

Reinforced variation and selection

Long-Evans rats were reinforced for generating variable sequences of four left (L) and right (R) leverpress responses. If the current sequence, for example, LLRR, differed from each of the preceding five sequences, then a food pellet was provided. Otherwise, there was a brief time-out. After the rats were generating a variety of sequences, one sequence was concurrently reinforced whenever it occurred, that is, an “always-reinforced” contingency was superimposed. The frequency of that sequence increased significantly from the variable baseline (Experiment 1). The “difficulty” of the always-reinforced sequence influenced the extent to which its frequency increased—“easy” sequences increased greatly, but “difficult” ones increased not at all (Experiment 2, and a replication with pigeons in Experiment 3). However, when reinforcement for baseline variability was systematically decreased while a difficult sequence was always reinforced, that sequence slowly emerged and attained relatively high levels (Experiments 4 and 5). Thus, a technique for training even highly improbable, or difficult-to-learn, behaviors is provided by concurrent reinforcement of variation and selection.     

Choice responding following multiple schedule training

Pigeons’ choice responding on 10-sec interpolated probes was studied after baseline training on multiple variable-interval variable-interval schedules of food reinforcement. Unreinforced choice following training with three different relative reinforcement rates (Experiment 1), with a 3-ply multiple schedule (Experiment 2), and with three different relative reinforcement durations (Experiment 3) was examined. Least squares lines were fit to choice relative response rate and schedule relative response rate as functions of training relative reinforcement rate; choice slope was significantly greater than schedule slope in all three experiments. This result is counter to the prediction of Herrnstein’s (1970) theory that these slopes should not differ. Luce’s (1959) theory also failed to account for the data. It was concluded that choice responding was controlled by both approach to the stimulus associated with the smaller mean interreinforcer interval or the longer duration, and avoidance of the other stimulus.     

The role of response-reinforcer correlation in signaled reinforcement effects

In three experiments, we examined the effects of signaling reinforcement during operant responding in order to illuminate the factors underlying instrumental overshadowing and potentiation effects. Specifically, we examined whether signaling reinforcement produces an enhancement and attentuation of responding when the response-reinforcer correlation is weak and strong, respectively. In Experiment 1, rats responded on variable-ratio (VR) or variable-interval (VI) schedules that were equated for the number of responses emitted per reinforcer. A signal correlated with reinforcement enhanced response rates on the VR schedule, but attenuated response rates were produced by the signal on the VI schedule. In Experiment 2, two groups of rats responded on a VI schedule while the two other groups received a conjoint VI, negative fixed-ratio schedule in which the subjects lost the availability of reinforcements if they emitted high response rates. A reinforcement signal attenuated responding for the simple VI groups but not for the animals given the negative fixed-ratio component, although the signal improved response efficiency in both groups. In Experiment 3, a poor correlation between responding and reinforcement was produced by a VI schedule onto which the delivery of response-independent food was superimposed. A signal for reinforcement initially elevated responding on this schedule, relative to an unsignaled condition; however, this pattern was reversed with further training. In sum, the present experiments provide little support for the view that signaling reinforcement enhances responding when the response-reinforcer correlation is weak and attenuates responding when this correlation is strong.     

Additive summation by stimulus compounding irrespective of behavioral contrast during discrimination training: An investigation with positive reinforcement and avoidance schedules

The similarity in the discrimination training leading to behavioral contrast and that preceding tests producing response enhancement to combined discriminative stimuli suggested that the two phenomena might be related. This was investigated by determining if contrast indiscrimination training was necessary for this outcome of stimulus compounding. Responding to tone, light, and to the simultaneous absence of tone and light (T + L) was maintained during baseline training by food reinforcement in Experiment I and by shock avoidance in Experiment II. During subsequent discrimination training, responding was reduced in T + L by programming nonreinforcement in Experiment I and safety or response-punishment in Experiment II. In the first experiment, one rat exhibited positive behavioral contrast, i.e., tone and light rates increased while his T + L rate decreased. In Experiment II, rats punished in T + L showed contrast in tone and light, this being the first demonstration of punishment contrast on an avoidance baseline with rats. The discrimination acquisition data are discussed in the light of current explanations of contrast by Gamzu and Schwartz (1973) and Terrace (1972). During stimulus compounding tests, all subjects in both experiments emitted more responses to tone-plus-light than to tone or light (additive summation). An analysis of the terminal training baselines suggests that the factors producing these test results seem unrelated to whether or not contrast occurred during discrimination training. It was concluded that the stimulus compounding test reveals the operation of the terminal baseline response associations and reinforcement associations conditioned on these multicomponent free-operant schedules of reinforcement.

     

Overshadowing by food of stimulus control by a keylight in a two-element serial compound

Pigeons were given successive discrimination training in which pecking during a choice period when the key was white was either reinforced or not, depending upon the prior presence or absence of a discriminative stimulus, which was a two-element serial compound. The compound consisted of a keylight and food, with food presented second or first in a forward or backward pairing for different groups of pigeons. In Experiment 1, the sequence was an S+ indicating reinforced trials, while in Experiment 2, the sequence was an S? indicating nonreinforced trials. Following acquisition of discriminated operant behavior, a sequence generalization test was administered during which all possible orders of the two stimuli were presented on test trials prior to the onset of the choice period. The results showed that food overshadowed stimulus control by the color of the light on the key on the sequence-generalization test, independently of whether food was presented first or second during training and independently of whether food was associated with reinforcement or nonreinforcement. The similarity of results for the two experiments suggests that overshadowing occurs independently of whether the compound is a discriminative stimulus for reinforcement or nonreinforcement. Simultaneous presentation of elements of a compound stimulus is not necessary for overshadowing because the phenomenon was captured with sequentially presented stimuli.     

Instrumental escape learning in the rat at 24-hour intertriai interval: Effects of magnitude and schedule of reinforcement

Three experiments investigated the effects of magnitude and schedule of reinforcement and level of training in instrumental escape learning at a 24-h intertriai interval. In Experiment I, two magnitudes of reinforcement were factorially combined with two schedules of reinforcement (CRF and PRF). Under PRF, large reward produced greater resistance to extinction than did small reward, while the reverse was true under CRF. In Experiment II, two levels of acquisition training were factorially combined with three schedules of reinforcement (CRF, single-alternation, and nonalternated PRF). Patterned running was observed late in acquisition in the single-alternation extended-training condition. Resistance to extinction was greater for the nonalternated PRF condition than for the single-alternation condition following extended acquisition, and the reverse was true following limited acquisition. Experiment III confirmed the extinction findings of Experiment II. The results of all three experiments supported an analysis of escape learning at spaced trials in terms of Capaldi’s (1967) sequential theory.     

Resistance to extinction as a function of percentage of reward: A reinforcement-level interpretation

Three experiments investigated the effects of percentage of reinforcement on the resistance to extinction of an instrumental running response. In Experiment 1, with N-length held constant, 47% reinforcement during acquisition generated greater resistance to extinction (R_n) than did 77%. In Experiment 2, this result was replicated with both functional N-length and number of N-R transitions held constant. In Experiment 3, R_n was shown to be a function of both N-length and percentage of reinforcement. The results of all three experiments were discussed in terms of Capaldi’s reinforcement level theory and possible alternative explanations.     

The effect of noncontingent outcomes on extinction of the response-outcome association

The effect of noncontingent outcomes on an instrumental response-outcome (R-O) association was examined in four experiments using transfer tests. In each experiment, rats were first given instrumental discrimination training designed to establish different stimuli as signals (S+s) for different outcomes. Transfer responses were subjected to different treatments across the experiments and then tested with the S+s. In Experiments 1 and 2, two transfer responses were both initially trained with two contingent outcomes. Then, each transfer response was subjected either to the addition of noncontingent presentations of one of those outcomes (Experiment 1) or to the replacement of one of the contingent outcomes with noncontingent presentations of that outcome (Experiment 2). Transfer tests revealed no significant difference in the ability of an S+ to promote performance of a transfer response based on their shared association with either the contingent or the noncontingent outcome. These results suggest that a response reinforced with two outcomes remains equally well associated with both of those outcomes despite prolonged exposure to noncontingent presentations of one of those outcomes. In Experiments 3 and 4, the possibility that the noncontingent schedules of reinforcement used in Experiments 1 and 2 might be capable of establishing an association between a response and its noncontingent outcome was examined. Transfer responses were trained with one contingent outcome and a different noncontingent outcome. Performance of these transfer responses was augmented more by presentations of an S+ trained with the contingent outcome than with the noncontingent outcome. These results confirm previous reports that instrumental responses are sensitive to outcome contingencies in acquisition and that noncontingent outcome presentations do not weaken previously established R-O associations. Several explanations are considered for the failure of subsequent noncontingent presentations of an outcome to reduce the strength of its association with the instrumental response.     

Pigeons’ use of cues in a repeated five-trial-sequence, single-reversal task

We studied behavioral flexibility, or the ability to modify one’s behavior in accordance with the changing environment, in pigeons using a reversal-learning paradigm. In two experiments, each session consisted of a series of five-trial sequences involving a simple simultaneous color discrimination in which a reversal could occur during each sequence. The ideal strategy would be to start each sequence with a choice of S1 (the first correct stimulus) until it was no longer correct, and then to switch to S2 (the second correct stimulus), thus utilizing cues provided by local reinforcement (feedback from the preceding trial). In both experiments, subjects showed little evidence of using local reinforcement cues, but instead used the mean probabilities of reinforcement for S1 and S2 on each trial within each sequence. That is, subjects showed remarkably similar behavior, regardless of where (or, in Exp. 2, whether) a reversal occurred during a given sequence. Therefore, subjects appeared to be relatively insensitive to the consequences of responses (local feedback) and were not able to maximize reinforcement. The fact that pigeons did not use the more optimal feedback afforded by recent reinforcement contingencies to maximize their reinforcement has implications for their use of flexible response strategies under reversal-learning conditions.     

“counting” by pigeons: Discrimination of the number of biologically relevant sequential events

Numerical competence has been studied in animals under a variety of conditions, but only a few experiments have reported animals’ ability to detect absolute number. Capaldi and Miller (1988) tested rats’ ability to detect absolute number by using biologically important events—the number of reinforced runs followed by a nonreinforced run—and found that the rats ran significantly slower on the nonreinforced run. In the present experiments, we used a similar procedure. Pigeons were given a sequence of trials in which responding on the first three trials ended in reinforcement but responding on the fourth trial did not (RRRN). When the response requirement on each trial was a single peck (Experiment 1), we found no significant increase in latency to peck on the fourth trial. When the response requirement was increased to 10 pecks (Experiment 2), however, the time to complete the peck requirement was significantly longer on the nonreinforced trial than on the reinforced trials. Tests for control by time, number of responses, and amount of food consumed indicated that the pigeons were using primarily the number of reinforcements obtained in each sequence as a cue for nonreinforcement. This procedure represents a sensitive and efficient method for studying numerical competence in animals.     

History effects on induced and operant variability

Two experiments evaluated history effects on induced and operant variability. College students typed three-digit sequences on a computer keyboard. Sequence variability was induced (by no reinforcement or variation- independent reinforcement) or reinforced (by variation- or repetition-dependent reinforcement). Conditions with induced and operant variability were presented according to a reverse between-groups design. In Experiment 1, we examined transitions from the variation or repetition contingencies to no reinforcement, and vice versa. In Experiment 2, the variation or repetition contingencies were followed or preceded by variation-independent reinforcement. The results showed that (1) a history of no reinforcement impaired operant variability learning; (2) induced variability levels were higher and lower after a history of reinforcement for variation and repetition, respectively; (3) repetition was more easily disrupted by no reinforcement and independent reinforcement than was variation; and (4) response variability and stability were a function of past and current reinforcement conditions. These results indicate that reinforcement history influences both induced and operant variability levels.     

Operant variability when reinforcement is delayed

Operant responses are often weakened when delays are imposed between the responses and reinforcers. We examined what happens when delayed reinforcers were contingent upon operant responsevariability. Three groups of rats were rewarded for varying their response sequences, with one group rewarded for high variability, another for middle, and the third for low levels. Consistent with many reports in the literature, responding slowed significantly in all groups as delays were lengthened. Consistent with other reports, large differences in variability were maintained across the three groups despite the delays. Reinforced variability appears to be relatively immune to disruption by such things as delays, response slowing, prefeeding, and noncontingent reinforcement. Furthermore, the small effects on variability depended on baseline levels: As delays lengthened, variability increased in the low group, was statistically unchanged in the middle group, and decreased in the high group, an interaction similar to that reported previously when reinforcement frequencies were lowered. Thus, variable operant responding is controlled by reinforcement contingencies, but sometimes differently than more commonly studied repetitive responding.     

Acquisition with partial and continuous reinforcement in pigeon autoshaping

Contemporary time accumulation models make the unique prediction that acquisition of a conditioned response will be equally rapid with partial and continuous reinforcement, if the time between conditioned stimuli is held constant. To investigate this, acquisition of conditioned responding was examined in pigeon autoshaping under conditions of 100% and 25% reinforcement, holding intertrial interval constant. Contrary to what was predicted, evidence for slowed acquisition in partially reinforced animals was observed with several response measures. However, asymptotic performance was superior with 25% reinforcement. A switching of reinforcement contingencies after initial acquisition did not immediately affect responding. After further sessions, partial reinforcement augmented responding, whereas continuous reinforcement did not, irrespective of an animal's reinforcement history. Subsequent training with a novel stimulus maintained the response patterns. These acquisition results generally support associative, rather than time accumulation, accounts of conditioning.     

Learned variability

Potential contributors to sustained levels of variability in the topography of the rat’s barpress were investigated in two experiments. Behavior was classified into discretely defined components, and changes in components and their sequential organization were analyzed. Experiment 1 showed that topographic variability in the rat is modulated by shifts in reinforcement schedules. Variability decreased between either dipper training or extinction and continuous reinforcement (CRF), and increased between CRF and extinction. Once the press was acquired, variability did not change if the schedule (CRF) did not change. Experiment 2 showed that, regardless of subsequent changes in topographic requirements, rats initially shaped to press under more stringent criteria sustained higher levels of variability during CRF, but not during extinction, than rats shaped with less stringent criteria. The results suggest that subjects learn not only what to do during reinforcement but also how differently or variably to do it.     

Partial reinforcement in autoshaping with pigeons

The acquisition, maintenance, and extinction of autoshaped responding in pigeons were studied under partial and continuous reinforcement. Five values of probability of reinforcement, ranging from .1 to 1.0, were combined factorially with five values of intertrial interval ranging from 15 to 250 sec for different groups. The number of trials required before autoshaped responding emerged varied inversely with the duration of the intertriai interval and probability of reinforcment, but partial reinforcement did not increase the number of reinforcers before acquisition. During maintained training, partial reinforcement increased the overall rate of responding. A temporal gradient of accelerated responding over the trial duration emerged during maintenance training for partial reinforcement groups, and was evident for all groups in extinction. Partial reinforcement groups responded more than continuous reinforcement groups over an equivalent number of trials in extinction. However, this partial-reinforcment extinction effect disappeared when examined in terms of the omission of “expected” reinforcers.