Date of Award
Master of Science (MS)
We build off of previous ideas used to study both reductions between CSPrefutation problems and improper learning and between CSP-refutation problems themselves to expand some hardness results that depend on the assumption that refuting random CSP instances are hard for certain choices of predicates (like k-SAT). First, we are able argue the hardness of the fundamental problem of learning conjunctions in a one-sided PAC-esque learning model that has appeared in several forms over the years. In this model we focus on producing a hypothesis that foremost guarantees a small false-positive rate while minimizing the false-negative rate for such hypotheses. Further, we formalize a notion of CSP-refutation reductions and CSP-refutation completeness that and use these, along with candidate CSP-refutatation complete predicates, to provide further evidence for the hardness of several problems.
Brendan Juba Jeremy Buhler William Yeoh