Abstract:
|
Bayesian optimization is an effective tool for tuning the parameters of machine learning systems via A/B tests. However, the ability to jointly optimize many parameters of the system policy can be limited by the low throughput of online field experiments. We describe how a simple, biased, offline simulator can be used to accelerate the optimization by learning a multi-task model that combines observations from both online and offline tests. We show results from multi-task Bayesian optimization of a Facebook ranking system, where augmenting the online tests with the simulator allows for jointly optimizing up to 20 parameters with as few as 40 total A/B tests. Finally, we analyze factors behind model generalization and identify settings where multi-task Bayesian optimization is most beneficial.
|