This work makes significant progress on a relatively natural theoretical problem of proper agnostic learning halfspaces with a margin. The paper is technically strong and is a clear accept. A downside is that the problem and proposed algorithms appear to be largely irrelevant to practice.