Artificial grammar learning (AGL) has been used extensively to study implicit learning. In this task participants first observe letter sequences generated by a grammar. In a later test phase participants are asked to distinguish new grammatical and ungrammatical sequences. Participants are able to do this, both when the letters instantiating the grammar remain the same (standard AGL) and when the letters are changed between training and test (transfer AGL).

Virtually all models of AGL assume that there is no learning during the test phase. Yet, test learning can occur in AGL and the structural constraints of a grammar can imply useful cues at test as well as at training. For example, grammatical test sequences are often more similar to each other than are ungrammatical test sequences to each other. Similarity to test sequences observed so far can then be used as a cue for classification.

In the current research I used an episodic memory model, Minerva II, in order to simulate a recent study by Hendricks et al. (2013). They found that for standard AGL performing dual tasks at test was more detrimental to performance than dual tasks at training. For transfer AGL performing dual tasks at training reduced performance as much as dual tasks at test. The authors interpreted these results as revealing automatic vs. intentional process in AGL: transfer AGL requires intentional processes at both training and test, whereas standard AGL requires intentional processes at test but only automatic processes at training.

I modelled these experiments using a version of Minerva II extended to learn at test. The model encodes sequences probabilistically into memory based on a learning rate at both training and test. Each test sequence is classified based on the similarity to sequences encoded in memory so far, so that test sequences also influence classification. The model does not distinguish between automatic and intentional processes. The learning rate at training was varied independently of the learning rate at test in order to simulate dual task manipulations in different phases of the task. In order to model transfer AGL I used a simple repetition coding scheme in Minerva II.

For standard AGL the simulations revealed that learning rate at test had a much greater impact on classification than learning rate at training in Minerva II.  In contrast, for transfer AGL the effects of changing learning rates at training was the same as changing learning rate during test. In essence, the empirical data may not reveal automatic vs. intentional processes, but simply effects of a single similarity process. The simulation results and the notion of test learning invites useful avenues for further computational and empirical research in order to establish the processes involved in implicit learning.