GP-UCB
An acquisition function in Bayesian optimization that balances exploration and exploitation by adding the posterior mean to a confidence coefficient multiplied by the posterior standard deviation. The coefficient controls the strength of exploration.
Timeline (last 90 days)
GP-UCB selects the point with the largest posterior mean plus β times the posterior standard deviation as the next exploration target, where β controls the aggressiveness of exploration, and this strategy theoretically guarantees sublinear regret.
The paper does not disclose the specific efficiency improvement of GP-UCB over ε-greedy.
For resource allocation-type sequential decision problems, practitioners should exploit the structural similarity between actions rather than treating each action independently, and should use the Matérn + Linear composite kernel as a default starting point, calibrating it in the simulator for the specific environment before deployment.
GP-UCB uses a Gaussian process prior to model the reward function and defines the correlation between adjacent configurations through a kernel function, thereby sharing information across the configuration space and reducing exploration cycles.
The research team conducted a large-scale randomized controlled trial (RCT) in a real production environment, comparing the GP-UCB method with the standard ε-greedy baseline, and the results show that GP-UCB is statistically significantly better than the baseline.
The cumulative regret upper bound of standard GP-UCB is parameterized by the RKHS norm bound and the maximum information gain.
Standard Bayesian optimization acquisition functions (such as GP-UCB, EI, Thompson sampling) are almost all designed to minimize cumulative regret, and therefore naturally favor exploitation over exploration.
All Facts (7)
GP-UCB selects the point with the largest posterior mean plus β times the posterior standard deviation as the next exploration target, where β controls the aggressiveness of exploration, and this strategy theoretically guarantees sublinear regret.
50%UnverifiedThe paper does not disclose the specific efficiency improvement of GP-UCB over ε-greedy.
50%UnverifiedFor resource allocation-type sequential decision problems, practitioners should exploit the structural similarity between actions rather than treating each action independently, and should use the Matérn + Linear composite kernel as a default starting point, calibrating it in the simulator for the specific environment before deployment.
50%UnverifiedGP-UCB uses a Gaussian process prior to model the reward function and defines the correlation between adjacent configurations through a kernel function, thereby sharing information across the configuration space and reducing exploration cycles.
50%UnverifiedThe research team conducted a large-scale randomized controlled trial (RCT) in a real production environment, comparing the GP-UCB method with the standard ε-greedy baseline, and the results show that GP-UCB is statistically significantly better than the baseline.
50%UnverifiedThe cumulative regret upper bound of standard GP-UCB is parameterized by the RKHS norm bound and the maximum information gain.
50%UnverifiedStandard Bayesian optimization acquisition functions (such as GP-UCB, EI, Thompson sampling) are almost all designed to minimize cumulative regret, and therefore naturally favor exploitation over exploration.
50%