Senior Research Scientist
Branislav Kveton is a machine learning scientist at Adobe Research in San Jose. He was at Technicolor’s Research Center from 2011 to 2014, and at Intel Research from 2006 to 2011. Before 2006, he was a graduate student in the Intelligent Systems Program at the University of Pittsburgh. His advisor was Milos Hauskrecht.
He proposes, analyzes, and applies algorithms that learn incrementally, run in real time, and converge to near optimal solutions as the number of training examples increases. Most of his recent work is focused on online learning of structured problems, such as graphs, submodularity, matroids, polymatroids, and reinforcement learning.
Practical problems are often so massive that even low-order polynomial-time solutions are not practical. Fortunately, many optimization problems can be solved greedily, either optimally or suboptimally with guarantees. Two popular examples of such problems are finding the maximum of a modular function on a matroid and finding the maximum of a submodular function subject to a cardinality constraint. Recently, he proposed several algorithms for solving this kind of problems when the model of the problem is initially unknown / imperfect, and is learned by interacting repeatedly with the environment. These algorithms can solve many interesting real-world problems, such as learning near-optimal preference elicitation policies from eliciting preferences, and learning optimal policies for network routing from repeated rerouting.