Workshop: Chris Towell and Konrad Rawlik

Chris Towell

An experiment to assess the consistency of redundancy resolution during planar tasks

My previous work has focused on learning redundancy resolution policies from data where the policy is only partially observable by combining observations from different tasks. Central to this approach is the assumption that, for a range of tasks, humans use a single underlying redundancy resolution policy. Is such an assumption valid? While motor control theories which seek to minimize some quantity would support this assumption, other work suggests that there is a large amount of variability in how redundancy is resolved. There is currently no work comparing redundancy resolution strategies between different tasks. Another key question is: at what level is the redundancy resolution policy? Some studies suggesting high variability look only at the positional level, but might a velocity or acceleration level be more appropriate? I will describe my ongoing experiment to examine these questions.

Konrad Rawlik

A kernel based approach to Stochastic Optimal Control

Recent work on stochastic optimal control has demonstrated how the solution to the problem can be formulated in terms of path expectations without requiring explicit minimisation. This makes it possible to estimate the value function using either Monte Carlo methods or standard variational approximations. However both of these approaches suffer from significant problems in practical applications. I will discuss these problems and introduce a kernel based estimation framework which alleviates them.