Solving Expensive Optimization Problems in Dynamic Environments With Meta-Learning.

Area of Science:

• Computational Science
• Artificial Intelligence
• Optimization Theory

Background:

• Dynamic environments present significant challenges for expensive optimization problems due to time-varying objective functions.
• Existing data-driven evolutionary optimization and Bayesian optimization (BO) methods are less explored in dynamic settings.

Purpose of the Study:

• To propose a flexible meta-learning-based optimization framework for expensive dynamic optimization problems.
• To enable efficient adaptation to changing environments within limited computational budgets.

Main Methods:

• A meta-learning component uses gradient-based meta-learning to acquire experience (model parameters) across varying dynamics.
• An adaptation component utilizes learned parameters as initializations for rapid, few-shot adaptation in dynamic environments.

Main Results:

• The proposed framework demonstrates effectiveness on benchmark problems with diverse dynamic characteristics.
• Experimental results show superior performance compared to several state-of-the-art algorithms.

Conclusions:

• The meta-learning framework offers a simple yet effective solution for expensive dynamic optimization.
• It facilitates quick initialization of search in new environments, crucial for computational efficiency.