Explainable AI (XAI) for Spectral Analysis via Reinforcement Learning: Learning to Optimize

Spectral analysis is extensively utilized in various industrial and academic domains to examine the characteristics of unknown samples. On one hand, classic methods usually construct complex physical models to mathematically formulate the problem and then exploit suitable optimizers to iteratively converge to a solution. While effective, these approaches often suffer from long computational time and convergence problems at high dimensions. On the other hand, artificial intelligence (AI), particularly deep learning-based methods, has recently been proposed to address such issues and proven to be fast, efficient, and accurate. However, the black-box nature of neural networks often raises concerns among domain experts regarding the interpretability and trustworthiness of the results. To overcome these challenges, we propose a reinforcement learning-based framework as an explainable AI (XAI) approach for spectral analysis. Instead of directly generating an end-to-end (E2E) solution from input data using neural networks, we train an agent to serve as an intelligent optimizer. During the optimization iterations, the agent observes the environment and makes sequential decisions to update parameters toward convergence. ... mehrThis process reveals the intermediate steps and the reasoning behind the optimization, making it both explainable and verifiable. Experimental results on the benchmark dataset demonstrate that our method consistently achieves faster convergence than other popular classical optimizers, delivering accurate results with fewer iterations. These properties make the framework a promising tool for real-world deployment in industrial and scientific applications.