Offline Optimization

Offline optimization tackles the challenge of maximizing an unknown objective function using only a pre-existing dataset, crucial for expensive or dangerous real-world applications like materials design and drug discovery. Current research emphasizes developing robust surrogate models and efficient search strategies, including policy-guided gradient search and generative models, to overcome limitations of relying solely on learned function approximations. This field is actively developing benchmarks and algorithms to improve the accuracy and efficiency of offline optimization across single and multi-objective problems, impacting diverse fields requiring data-driven decision-making under constraints. The development of projection-free methods and efficient algorithms for large-scale problems are also key areas of focus.