Abstract:
The brain is really good at controlling movement. Consequently, theories of motor function often assume that movement is controlled in the best way possible - i.e. optimally. Indeed, models derived from optimal control theory have provided the most parsimonious and accurate accounts of many empirical phenomena. While traditional emphasis has been on average trajectories and open-loop control sequences, we have recently redefined optimality on the level of feedback control laws and focused on the sensory-motor loops that generate complex behavior online. This has made it possible to fit a number of previously unrelated observations - such as motor synergies, controlled parameters, goal-directed corrections, structured variability, speed-accuracy tradeoffs, kinematic regularities, eye-hand coordination patterns - into what may become a unified theoretical framework for interpreting motor function. In my talk I will illustrate the richness of this framework. The main focus will be on motor coordination - i.e. the simultaneous control of redundant actuators towards the achievement of a common goal. I will show that optimal feedback control leads to a "minimal intervention" principle, where trial-to-trial deviations from the average trajectory are corrected only if they interfere with the task goals. As a consequence, the optimal controller uses only a subspace of the available control space, and allows variability to accumulate in redundant dimensions. This explains one of the most puzzling characteristics of biological movements: they can be extremely variable, and yet always accomplish their goal. I will also present efficient new methods for iterative optimal controller design, which generalize the Linear-Quadratic-Gaussian methodology to control- and state-dependent noise, as well as to constrained nonlinear systems subject to non-quadratic costs. These methods are likely to extend the applicability of optimal control theory in the field of biological movement, and may be of independent interest in other fields as well.