Inverse problems are fundamental to a variety of scientific and engineering problems, but are commonly ill-posed, nonlinear, and sensitive to measurement errors. In this study, a novel machine learning based ABC framework (ML-ABC) is introduced to enhance the stability of and the efficiency in IP estimation. The proposed strategy includes a surrogate learning, which is used to mimic the objective landscape and guide the injection of candidates during exploration phase, thus achieving faster convergence rate and avoided premature stagnation than those of the conventional ABC algorithm. The method is tested on inverse reconstruction problems in simulations, where unknown model parameters are estimated from noisy observations. Experimental results illustrate that ML-ABC achieves higher reconstruction quality and a more consistent estimate over multiple levels of noise compared to previous methods while having the ability to perform global search. Further, our proposed approach produces interpretable convergence behavior and can be used for error-surface analysis that could help in verifying the reliability of solutions in ill-posed conditions. In summary, the ML-ABC is a feasible hybrid optimization framework to address inverse problems, providing better robustness and computational efficiency.
Al-thabhawee et al. (Sat,) studied this question.