Inference Regret
A metric that measures the gap between the recommendation point ultimately returned by an optimization algorithm and the global optimum. It focuses on the quality of the final output rather than the cumulative loss incurred during the process.
Timeline (last 90 days)
This study proposes that the existing cumulative regret guarantees of Bayesian optimization do not necessarily characterize the quality of the algorithm's final output.
Inference regret measures the gap between the global optimum value and the function value of the point the algorithm finally returns.
The researchers bound the upper bound of the Gaussian process posterior standard deviation over the entire search space via the filling distance of the queries, and thereby derive an upper bound on the inference regret.
The tightness and applicability of this theoretical guarantee on non-stationary or high-dimensional problems still require more follow-up research to verify.
All Facts (4)
This study proposes that the existing cumulative regret guarantees of Bayesian optimization do not necessarily characterize the quality of the algorithm's final output.
50%UnverifiedInference regret measures the gap between the global optimum value and the function value of the point the algorithm finally returns.
50%UnverifiedThe researchers bound the upper bound of the Gaussian process posterior standard deviation over the entire search space via the filling distance of the queries, and thereby derive an upper bound on the inference regret.
50%UnverifiedThe tightness and applicability of this theoretical guarantee on non-stationary or high-dimensional problems still require more follow-up research to verify.
50%