Core Concepts
Primal-dual algorithms with weakly adaptive regret minimizers can optimize bidding in online ad auctions under budget and ROI constraints.
Abstract
The paper addresses online learning problems with costly decisions.
Introduces weakly adaptive regret minimizers to handle budget and ROI constraints.
Shows how to optimize bidding in various practical mechanisms.
Guarantees no-regret outcomes under stochastic and adversarial inputs.
Provides a framework for repeated non-truthful auctions.
Stats
First, the decision maker must know the value of parameters related to strict feasibility.
Second, a strictly feasible solution must exist at each round.
ROI constraints are not packing, complicating the application of known algorithms.
Quotes
"We show how such assumptions can be circumvented by endowing standard primal-dual templates with weakly adaptive regret minimizers."
"This results in a 'dual-balancing' framework which ensures that dual variables stay sufficiently small."