Trustworthy and Sustainable Automate Design of Algorithms for the Generative AI Era

Objective

Algorithms are the key to solving complex problems across diverse research and application domains. However, traditional algorithm design relies heavily on human expertise, making algorithm design a costly and time-consuming process. While recent advances in Algorithm Configuration (AC) have partially automated the design of algorithms, their success still strongly relies on human-crafted algorithmic frameworks, limiting broader adoption across domains. With the advent of Large Language Models (LLMs), a new paradigm for Automated Design of Algorithms (ADA) has emerged, enabling the direct evolution of algorithm implementation from a significantly broader set of candidate operators, eliminating the need for handcrafted frameworks. Despite these advances, current ADA research primarily focuses on achieving competitive performance while neglecting trustworthiness, explainability, and sustainability, which are critical for practical deployment. This project, TSADA, aims to address these issues by developing methods to enhance the explainability, transparency, and efficiency of ADA, including both traditional AC and LLM-driven approaches. Through working on the design of Satisfiability (SAT) solvers, I will propose sustainable ADA techniques capable of leveraging knowledge of prior algorithms to minimize redundant training tasks. Furthermore, I will systematically compare traditional AC and LLM-driven ADA methodologies to provide evidence-based guidelines for selecting the most suitable approach under different scenarios. By advancing explainable and resource-efficient ADA, TSADA contributes to the broader objectives of Explainable AI and Green AI, paving the way for more trustworthy, transparent, and sustainable algorithmic innovation.