Instance Generation for Meta-Black-Box Optimization through Latent Space Reverse Engineering

The recent introduction of LSRE (Latent Space Reverse Engineering) marks a significant advancement in Meta-Black-Box Optimization (MetaBBO). Traditional training methods often rely on the CoCo-BBOB benchmark suite, which lacks diversity and poses a risk of overfitting, ultimately hindering the generalization capabilities of optimization algorithms. LSRE addresses this issue by training an autoencoder to map high-dimensional problem features into a two-dimensional latent space, allowing for the generation of a more diverse set of training problems, termed Diverse-BBO. This innovation not only enhances the training process for various MetaBBOs but also promises to improve the adaptability of low-level optimizers to new, unseen problem instances. As the field of automated algorithm design evolves, the effectiveness of LSRE could lead to more robust and versatile optimization solutions.