Personal tools
You are here: Home Events ANC/DTC Seminar: Andreas Krause, ETH Zurich (Host: Charles Sutton)

ANC/DTC Seminar: Andreas Krause, ETH Zurich (Host: Charles Sutton)

— filed under:

Learning to Optimize with Confidence

  • ANC/DTC Seminar
When Jan 29, 2013
from 11:00 AM to 12:00 PM
Where IF 2.33
Add event to calendar vCal

In many applications, we have to make decisions with unknown rewards.
Often, the number of choices is extremely large (possibly infinite).

We can model this problem as maximising an unknown reward function that is expensive evaluate from a small number of noisy samples.  A key challenge is trading off exploration (estimating the function) and exploitation (maximizing the estimated function).  In this talk, I will present recent progress towards this fundamental problem.  I will show how a simple confidence-guided sampling rule attains near-minimal regret for a large class of reward functions, modeled as samples from Gaussian processes or having low RKHS norm. I will further demonstrate how our approach allows to scale up through parallelization, incorporate context to more quickly solve related tasks, and address multi-objective tradeoffs. I will illustrate the approach in several real-world applications. Applied to experimental design for protein structure optimization, our approach enabled engineering of active P450 enzymes that are more thermostable than any previously made by chimeragenesis, rational design, or directed evolution.