How can agents collaborate even when they
disagree about the state of the world?
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
allows 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.
Below are recent papers
representing algorithms we have developed for a variety of
multi-agent optimization settings:
1. K. Yazdani and M. Hale, "Asynchronous Parallel Nonconvex Optimization Under the Polyak-Łojasiewicz Condition," Under review.
[preprint]
2. M. Blondin and M. Hale, "A Decentralized Multi-Objective Optimization Algorithm," Journal of Optimization Theory and Applications (JOTA). In press.
[link]
3. M. Ubl and M. Hale, "Totally asynchronous quadratic programming: Convergence rates, parameter selection, and tradeoffs," IEEE Transactions on Control
of Network Systems (TCNS). In press. [preprint]
4. K. Hendrickson and M.T. Hale, "Towards totally asynchronous primal-dual convex optimization in blocks,"
59th IEEE Conference on Decision and Control, pp. 3663-3668.
[link]
