Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)




The continuity-noncontinuity issue remains a focus for both the theoretical and experimental inquiry into the fundamental nature of the learning process. Although experimental studies of paired-associate learning have often disclosed continuities, mathematical models incorporating all-or-none processes have generally fit the data quite well. The most basic of these is the one-element model proposed by Bower in 1961. The model assumes only two states, a learned state and an unlearned state. Transitions from the unlearned to the learned state occur with a fixed probability which is constant across trials. Extensions and modifications of the Bower model have consisted basically of the addition of intermediate states which have their owned fixed transition probabilities. Theoretical explanations of these intermediate states include short-term memory stores, discrimination processes, and recognition-recall differences.

Previous evidence has shown that Bower's model loses its accuracy of prediction with difficult and/or long lists. In an effort to predict accurately both quickly and slowly learned lists an all-or- none three state model was built. The model, called the paired- associate recognition-recall (PARR) model, was based upon established differences between recognition and recall learning. The first state is a nonrecognition-nonrecall state in which the probability of a correct response is zero. The intermediate state or recognition state contains paired-associate items which can be recognized but are not yet recalled. While in the recognition state an item may be selected for rehearsal with probability p, in which case a correct response will be given. The third state is the recall or learned state in which pairs are correctly recalled on every trial. Unlike many previous models direct transitions from the first state to the third state are possible. Also, the probability of moving into the recall state from the recognition state is independent of the probability that an item is rehearsed. Predictions for the learning curves, errors before the first correct response (J), total errors (T), and last error trial (L) were derived and tested against obtained data. Predictions from the Bower model and from a model for discrete performance levels by Bower and Theios were also compared to the data.

List difficulty was varied by manipulating stimulus term meaningfulness. CVC’s selected from Archer's 1960 list were used to build low, medium, and high meaningfulness lists. Response terms were the digits 1-16 for each list.

None of the models tested adequately described data from the three meaningfulness conditions. In all cases the models predicted a more rapid rate of learning than was observed. The Bower-Theios and one-element models made very similar predictions about the learning curves but were very dissimilar in their predictions of the probability distributions of J, T, and L. Data from the high meaningfulness list indicated that an intermediate state did not exist. Since the Box^er- Theios and PARR models are intermediate state models only the one- element model was used in the consideration of the high meaningfulness data. Surprisingly, the one-element model provided a very bad fit to the data; predicting a much more rapid learning rate than was observed.

Results were discussed in terms of parameter estimates, the failure of the one-element model, and with regard to the conventional two-stage analysis of paired-associate learning.