Reinforcement-test intervals in paired-associate learning

Abstract Learning in paired-associate items was studied in relation to the lengths of the intervals between the reinforcements (Rs) and the test trials (Ts) on each item, in order to gather data on the roles of these intervals and of the Ts and Rs themselves in learning. The sequence of events for an item can be depicted as follows (where I denotes an intervening trial on another item): Some items had T 1 omitted entirely. The most salient aspects of the results were (a) the inclusion of T 1 enhanced performance on T 2 , (b) on items with T 1 omitted, as the R 1 –R 2 interval increased, the probability of being correct on T 2 ( P ( C 2 )) first increased and then declined, and (c) with the R 1 –R 2 interval length fixed, as T 1 moved from just after R 1 to just before R 2 , P ( C 2 ) rose and then fell. The nonmonotonicity of these results made it impossible to fit these data with any simple variant of stimulus sampling theory or extant Markov models with long- and short-term memory stores. A new model, with its main innovative feature the postulation of multiple short-term memory states, each with unique properties, is proposed; its fit to the data was quite satisfactory.