By contrast, there was little evidence that BOLD in the vmPFC was also related to value, as might have been expected, though parts of the medial temporal lobe also showed significant Romidepsin mouse correlations with these and with reward predictions. The authors suggested that the latter findings might possibly reflect
the spatial nature of the task, compared with the more abstract Markov decision problems that had previously implicated the vmPFC. Other regions including the anterior insula, the precentral cortex, and the medial cingulate covaried with facets of the transitions available from a room, suggesting that they might be involved in realizing the model of the world. Tasks such as this have a strong
spatial component, as opposed to the more abstract structure of many planning tasks, and this attribute might account for the presence of model-based GSK126 cell line signals seen in hippocampus and medial temporal lobe. In fact, this has a bearing on a suggestion that there is another form of controller, an “episodic controller,” that involves these very structures (Lengyel and Dayan, 2008). Other illuminating paradigms include a so-called grid-sailing task (Fermin et al., 2010), which uses structurally different rules (in the form of key mappings) in a motor-learning task. This task Ribonucleotide reductase has provided evidence that subjects use a model-based strategy to generalize learning. There is also a suggestion that explicit instructions and advice (whose immediate impact must surely be more model-based than model-free) operate by boosting the impact of model-free learning on trials on which instructions are followed (Doll
et al., 2009). Also, of note is a recent implementation of an ingenious behavioral design, involving a simple one- and two-step problem in which learning and performance occurred in distinct phases (Gershman et al., 2012). In the final phase of the task, model-based and model-free controllers would make the same choices, albeit for different reasons. In fact, the authors observed that subjects acted in a manner consistent with a model-based system having trained by a model-free one during an earlier phase of learning, as in an online or offline form of the DYNA-Q algorithms mentioned above (Sutton, 1991). In effect, these findings highlight cooperation, as opposed to competition, between the two systems. There are many outstanding questions related to model-based and model-free control, and these are now the focus of intense investigation. In the remainder of this Review, we touch on some of the main strands of this research and the plethora of unresolved issues.