Modelling the Detailed Pattern of SRT Sequence Learning

Any viable model of serial reaction time (SRT) sequence learning needs to be able to capture the relative extents to which participants learn different SRT sub-sequences. However, previous attempts to empirically establish a relative pattern of this sort have failed to fully control for ‘sequential-effects’, or have not satisfactorily ruled-out the influence of speed-accuracy trade-off. In this paper it is shown that, when sequential-effects are controlled for in a two-choice SRT task, participants learn those subsequences that end in an alternation better than those that end in a repetition. It is further demonstrated that, at least for the parameters investigated, a buffer network, a Jordan network, an SRN, an augmented SRN, and an AARN are unable to account for this pattern.