Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Time-Varying Gaussian Process Bandit Optimization

About

We consider the sequential Bayesian optimization problem with bandit feedback, adopting a formulation that allows for the reward function to vary with time. We model the reward function using a Gaussian process whose evolution obeys a simple Markov model. We introduce two natural extensions of the classical Gaussian process upper confidence bound (GP-UCB) algorithm. The first, R-GP-UCB, resets GP-UCB at regular intervals. The second, TV-GP-UCB, instead forgets about old data in a smooth fashion. Our main contribution comprises of novel regret bounds for these algorithms, providing an explicit characterization of the trade-off between the time horizon and the rate at which the function varies. We illustrate the performance of the algorithms on both synthetic and real data, and we find the gradual forgetting of TV-GP-UCB to perform favorably compared to the sharp resetting of R-GP-UCB. Moreover, both algorithms significantly outperform classical GP-UCB, since it treats stale and fresh data equally.

Ilija Bogunovic, Jonathan Scarlett, Volkan Cevher• 2016

Related benchmarks

TaskDatasetResultRank
Bayesian OptimizationHartmann d=6
Relative Batch Instantaneous Regret0.44
8
Black-box OptimizationLevy8 function
AUSR2.45e+3
8
Black-box OptimizationRosenbrock10 function
AUSR1.22e+3
8
Black-box OptimizationLangermann2 function
AUSR304.7
8
Black-box OptimizationGriewank 6 function
AUSR238.3
8
Financial Portfolio OptimizationPortfolio5 Asset Allocation
AUSR1.29e+3
8
Black-box OptimizationHartmann6
AUSR152.5
8
Robot Pushing TaskRobot4 Box2D simulation
AUSR50.6
8
Hyperparameter OptimizationUCI Breast Cancer MLP4
AUSR3.3
8
Bayesian OptimizationHartmann3 d+1=3
Average Regret0.26
6
Showing 10 of 20 rows

Other info

Follow for update