As discussed previously, there are a wide range of models that have been applied to the Wisconsin card sorting task. However, which one is most biologically plausible? That is, uses a realistic model of neurons, respects the constraints of the human brain and maps readily onto experimental data.
Amos (2000) and Monchi et al. (2000) use the similar approach of assigning each card attribute to a node and using mutual inhibition to choose the right one. Although their models are biologically plausible and make many neuroanatomical predictions, they are functionally implausible. Their networks are created for the unique purpose of of completing the WCST. This type of specialization isn't found in the brain as far as I know. I'm also having a really hard time understanding on Monchi is using Leaky-Integrate-Fire neurons to compute functions, but that might be a topic for another question.
Rougier et al. (2005) use Leabra, which has been criticized for it's lack of biological plausibility. For further elaboration on these criticism, which is mostly centered around the biologically-implausible use of a Winner-Takes-All computational block being used, check out How to Build a Brain by Chris Eliasmith.
Bishra et al. (2010) only define symbolic rules. There's no neuroanatomical mapping. Consequently, although they claim tuned parameters could predict performance, there doesn't seem to any definition of where these parameters come from.
In conclusion, although none of the approaches are perfectly biologically plausible, Amos and Monchi seem to definitely come out on top. Although hard-coded concepts such as the possible parameters for matching are totally acceptable, a model that integrates with a larger more general system while remaining biologically plausible would be ideal.
Rigotti et al. have a model of the wisconsin card sorting task using a neural network and compare it with data from prefrontal cortex http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2967380/