Two representative computational models for switching between internal models have been proposed (Imamizu et al., 2004), namely a mixture-of-experts model (Gomi and Kawato, 1993; Graybiel et al., 1994; Jordan and Jacobs, 1994), and a modular selection and identification for control model (MOSAIC) (Haruno et al., 2001; Doya et al., 2002; Wolpert et al., 2003). The behaviour in the mixture-of-experts model is determined by a gating module which selects between experts, while behaviour in the MOSAIC model combines recommendations from each expert. However, these two switching models alone seem unable to capture the essence of learned skills, learning behaviour and the occurrence of errors (cf. Diedrichsen et al., 2005; Thoroughman et al., 1999). In this poster, a novel hierarchical control architecture for switching internal models is proposed. The architecture, which is based on a combination of predictive and adaptive control, can account for the observed results on combined mixture-of-experts and MOSAIC control behaviour. Experimental trials on verbal fluency tasks are designed to tap the subjects switch costs (cf. Gurd et al., 2002). Subjects switch between producing verbal output from over-learned sequences, such as days of the week, months of the year, or letters of the alphabet (ie. Monday, January, A, Tuesday, February, B, etc. The added times the subjects take to switch between tasks, as well as the occurrence of errors, are measured. The results show that the deteriorating performance of the subjects (ie. as the number of categories are increased, or as the time into the task progresses), seems due to the inconsistency between predictive outputs in the internal models and actual outcomes on the performed tasks.
