Core Concepts
Bounce proposes a reliable algorithm for optimizing high-dimensional black-box functions in combinatorial and mixed spaces, outperforming state-of-the-art methods.
Abstract
This paper introduces Bounce, a novel algorithm for Bayesian optimization in high-dimensional spaces. It addresses the limitations of current methods by leveraging nested embeddings and trust region management. The experimental evaluation demonstrates Bounce's robust performance across various benchmarks, surpassing existing algorithms like BODi and COMBO. The algorithm's ability to handle mixed variable spaces efficiently makes it a valuable tool for real-world applications.
Abstract:
Impactful applications require optimizing high-dimensional black-box functions with mixed and combinatorial input spaces.
Current state-of-the-art methods are unreliable when the unknown optima lack structure.
Bounce proposes a novel algorithm that achieves reliable performance on high-dimensional problems.
Introduction:
Bayesian optimization is crucial for optimizing expensive-to-evaluate black-box functions with numerous applications.
Challenges arise from high-dimensionality and different variable types in real-world problems.
Recent efforts extend Bayesian optimization to combinatorial and mixed spaces.
Algorithm Overview:
Bounce uses nested embeddings to model GP in low-dimensional subspaces efficiently.
Trust region management allows focusing on promising regions of high-dimensional target spaces.
Batch parallelism enables parallel evaluations of the objective function, improving sample efficiency.
Experimental Evaluation:
Empirical evaluation shows Bounce outperforms state-of-the-art algorithms on diverse benchmarks.
Sensitivity analysis reveals shortcomings of existing methods like BODi and COMBO.
Bounce demonstrates robust performance across various tasks, making it a reliable optimizer for practitioners.
Conclusion:
Bounce offers a scalable solution for optimizing high-dimensional black-box functions in combinatorial and mixed spaces.
The algorithm's reliability and performance superiority make it a valuable tool for practical applications.
Stats
"BODi is more susceptible to the location of the optimizer than COMBO."
"COMBO's implementation suffers from a bug explained in Appendix H.2."
Quotes
"Bounce leverages parallel function evaluations efficiently."
"The proposed algorithm is reliable for high-dimensional black-box optimization."