Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks

The paper introduces Evolving Programmatic Bottlenecks (EPB), a novel framework for interpreting Neural Combinatorial Optimization (NCO) models by transforming them into human-readable program portfolios. EPB utilizes a large language model (LLM) to autonomously generate and refine programs, employing a hybrid textual-numerical gradient descent approach to optimize program capacity and performance. This advancement enhances the interpretability of NCO, facilitating better understanding and deployment of sequential decision-making models in practical applications.