Is There a Future for Genetic and Learning-Based Methods in Query Optimizers?

Over the last decades, many alternative approaches to query optimization have been proposed, including genetic algorithms, machine learning models, neural networks, and reinforcement learning techniques. These methods aim to address limitations of traditional dynamic programming, particularly for join ordering in large and complex queries. The goal of this talk is to critically analyze these approaches rather than promote a single solution. I review genetic algorithms, classical machine learning models, neural-network–based optimizers, and reinforcement learning techniques, discussing where they have shown promising results and where they consistently fail. In this talk, I compare reported results from the literature and, where possible, align experimental evaluations using standard benchmarks. The goal of the talk is to distinguish practical and reproducible techniques from purely research-oriented prototypes, and to clarify what role—if any—these methods can realistically play in industrial database query optimizers.