Communication/Computation Constrained Autonomy

This research focuses on finding theoretical guarantees for algorithms to converge under bounded and unbounded delays for convex and non-convex optimization problems.

As multi-agent systems grow, it becomes difficult or impossible to ensure that agents generate and share information with identical timing. Instead, agents may compute new information and communicate it with arbitrary timing. Doing so leads to asynchronous behaviors and, in particular, agents will receive different information at different times, causing them to have disagreeing information onboard. Despite these disagreements, agents must continue working together to complete some task.

Some multi-agent tasks in learning and control can be formalized as convex or nonconvex optimization problems. We are interested in solving both unconstrained and constrained problems in a multi-agent way that is robust to asynchrony. Of particular interest are scenarios in which delays can be arbitrarily long (rather than assuming delays obey a pre-specified bound on their length). Unbounded asynchrony can allow agents to have substantial disagreements about the values of decision variables in a problem, though careful algorithmic design can still provide strong, practical convergence guarantees.