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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Bayesian Optimization | Hartmann d=6 | Relative Batch Instantaneous Regret0.44 | 8 | |
| Black-box Optimization | Levy8 function | AUSR2.45e+3 | 8 | |
| Black-box Optimization | Rosenbrock10 function | AUSR1.22e+3 | 8 | |
| Black-box Optimization | Langermann2 function | AUSR304.7 | 8 | |
| Black-box Optimization | Griewank 6 function | AUSR238.3 | 8 | |
| Financial Portfolio Optimization | Portfolio5 Asset Allocation | AUSR1.29e+3 | 8 | |
| Black-box Optimization | Hartmann6 | AUSR152.5 | 8 | |
| Robot Pushing Task | Robot4 Box2D simulation | AUSR50.6 | 8 | |
| Hyperparameter Optimization | UCI Breast Cancer MLP4 | AUSR3.3 | 8 | |
| Bayesian Optimization | Hartmann3 d+1=3 | Average Regret0.26 | 6 |