UC San Diego

The transition from algorithmic to memory-based performance is a core component of cognitive skill learning. E.g., an arithmetic problem such as 4 x 3 may initially be solved with a repeated addition algorithm (4 + 4 + 4), but with practice the answer will be recalled directly from long-term memory (LTM). There has been debate about the temporal dynamics of strategy execution, with some models assuming a race (i.e., independent, capacity unconstrained parallel processing) between algorithm and retrieval, and others assuming a choice mechanism. This work introduces an original paradigm that permits (for the first time) the objective identification of strategy use on every trial, as well as the latency of each component step of an algorithm. I also introduce a technique for appropriately aggregating data across different learning curves. Results are uniquely consistent with a strategy choice mechanism involving a competition between the retrieval strategy and the 1st step of the algorithm. Some previously undiscovered skill-acquisition phenomena (such as increasing latency for algorithm initiation on trials immediately preceding the first correct direct retrieval for each item) are identified and discussed. Examination of partial-algorithm trials (in which the algorithm is initiated, but abandoned prior to completion in favor of direct retrieval) indicates that for algorithms consisting of multiple retrievals from LTM, the bottleneck extends beyond the 1st step of the algorithm, whereas for simple perceptual-motor algorithms, some parallel performance on later steps is possible. I introduce a theoretical framework that can accommodate the results found for different classes of algorithms. Results highlight the importance of studying partial-algorithm trials (something that has not been possible in previous skill-acquisition paradigms), and also the importance of considering the issue of efficiency in strategy scheduling as a factor that may affect performance over the course of practice