Abstract:Bayesian optimization is an effective technique for black-box optimization, but its applicability is typically limited to low-dimensional and small-budget problems due to the cubic complexity of computing the Gaussian process (GP) surrogate. While various approximate GP models have been employed to scale Bayesian optimization to larger sample sizes, most suffer from overly-smooth estimation and focus primarily on problems that allow for large online samples. In this work, we argue that Bayesian optimization algorithms with sparse GPs can more efficiently allocate their representational power to relevant regions of the search space. To achieve this, we propose focalized GP, which leverages a novel variational loss function to achieve stronger local prediction, as well as FocalBO, which hierarchically optimizes the focalized GP acquisition function over progressively smaller search spaces. Experimental results demonstrate that FocalBO can efficiently leverage large amounts of offline and online data to achieve state-of-the-art performance on robot morphology design and to control a 585-dimensional musculoskeletal system.
What problem does this paper attempt to address?
### What problems does this paper attempt to solve?
This paper aims to solve the computational efficiency and performance problems of Bayesian Optimization (BO) when dealing with high - dimensional problems and large - scale data sets. Specifically, the paper focuses on the following aspects:
1. **Computational complexity**: Traditional Bayesian optimization relies on the Gaussian Process (GP) as a surrogate model, but the computational complexity of the posterior covariance matrix of GP is \(O(n^3)\), which makes BO difficult to handle large - scale data sets or high - dimensional problems.
2. **Limitations of sparse Gaussian processes**: Although there are various sparse GP methods to expand the application range of BO, these methods are usually applicable to low - dimensional problems and perform poorly in high - dimensional settings. In addition, the classical sparse GP model tends to perform a smooth fitting of the entire function landscape, resulting in insufficient predictive ability in key regions.
3. **Efficient use of offline and online data**: Most existing methods only evaluate their performance on low - dimensional problems, and the effectiveness of existing methods is unclear for scenarios that include a large amount of offline data or high - dimensional problems.
To solve these problems, the paper proposes a new method - FocalBO, which improves Bayesian optimization in the following ways:
- **Focalized Gaussian Process (focalized GP)**: A new variational loss function is introduced, enabling the GP to fit the local function landscape more accurately, thereby improving the prediction accuracy in key regions.
- **Hierarchical sampling point selection strategy (hierarchical acquisition optimization)**: A hierarchical framework FocalBO is designed, which optimizes the selection of sampling points by gradually narrowing the search space, thus achieving a better balance between global and local.
Through these improvements, FocalBO can perform excellently when dealing with high - dimensional problems and large - scale data sets. Experimental results show that it has achieved state - of - the - art performance in tasks such as robot morphology design and 585 - dimensional musculoskeletal system control.
### Summary
The main contributions of the paper include:
1. Designing FocalBO, which adopts a hierarchical acquisition function optimization strategy to achieve efficient high - dimensional heterogeneous structure optimization.
2. Experimental results prove the superior performance of FocalBO in using large - scale offline data for online optimization, especially in the high - dimensional musculoskeletal system control problem, which involves the optimization of more than 500 variables.
Through these innovations, FocalBO significantly improves the applicability and performance of Bayesian optimization when dealing with large - scale data and high - dimensional problems.