This is a slightly disingenuous challenge because optimal control cannot reproduce handwriting as a result of requisite motion being solenoidal. As noted above, this is a shortcoming of optimal control when it comes to itinerant (sequential and wandering) movements. selleck chemicals In short, the compete class theorem suggests that any optimal trajectory specified by a cost function can be specified by a prior belief but that not every optimal trajectory can be specified by a cost function. The issues addressed in this review are largely theoretical in nature and speak to formal or computational modeling of motor control: specifically, should these models be

based on optimal control theory or optimal Bayesian inference. However, the answer

has some profound neurobiological implications. For example, if descending motor commands are top-down predictions, then descending motor efferents should share physiological and anatomical characteristics with top-down or backward connections in other systems. Indeed, descending projections from primary motor cortex share many features with backward connections in visual cortex: they originate in infragranular layers and target cells expressing NMDA receptors. This is somewhat paradoxical, from the orthodox perspective (Shipp, 2005), because backward modulatory characteristics (Sherman and Guillery, 1998) would not be expected of driving motor command signals. This apparent Selleck Trametinib paradox is resolved by active inference, which also provides a principled explanation for why the motor cortex is agranular (R. Adams, personal communication). There are clearly many operational issues that attend the distinction between optimal control and active inference. For example,

how does active inference compensate for altered limb dynamics or external perturbations? A treatment of this can be found in Friston et al. (2010), in which movement trajectories are shown to be remarkably robust to perturbations, very both to forces on a limb and fluctuations in motor gain. Heuristically, active inference counters unpredicted forces immediately (to suppress prediction errors on force); in contrast, optimal control can only adjust its (state-dependent) control signals after unpredicted forces change the state of the motor plant. Another key area we have not considered is the learning or acquisition of prior beliefs. In optimal control, the value function is learned, whereas in active inference, the problem reduces to learning the parameters (of the equations of motion) that constitute prior beliefs. This is a standard problem in inference and corresponds to perceptual learning. For example, the agent depicted in Figure 5 could optimize its parameters during action observation (with respect to free energy) and use them to reproduce observed behavior during action.