Login
Computer Science @ Rochester
Thursday, April 27, 2006
11:00 AM
CSB 209
Ph.D. Thesis Proposal
Manu Chhabra
University of Rochester
Optimality in Motor Control
Researchers have often proposed that stereotypical motor behavior is optimal. However, it is well known that finding the optimal solution to a motor task is computationally difficult. How is the motor system able to solve such a hard problem reliably and repeatedly? In Part 1 of the talk, we propose that this problem can be solved by composing the optimal solution as a linearsum of asynchronously activated components (or "motor synergies"). Using a simulated arm and a reaching task, we compute optimal control strategies and extract optimal synergies via a dimensionality-reduction technique. We show that optimal control strategies can be closely approximated as sums of these synergies, and that novel tasks can be rapidly solved by learning to linearly combine these synergies. In Part 2 of the talk, we test whether human subjects show optimal motor performance in unnatural environments. Subjects were asked to control a simulated object by selecting motor commands using a computer mouse so that the object spends as much time as possible in a target region. The dynamics between the motor commands and the object were corrupted with noise. We tested subjects under 3 conditions---no noise, proportional noise, and inversely proportional noise. Subjects quickly learned optimal control policies (compared to the information-theoretic baseline) in all three conditions. Further analysis showed that subjects were not using a fixed error-correction strategy--- their strategy was tailor-made for the specific noise condition. This result demonstrates that the human motor system is flexible enough to construct optimal control policies by using information about the noise condition. We conclude the talk with a summary of directions for future work.