Annotation

Annotation: Objectives. This paper examines the problem of tuning information retrieval results reranking models as a mixed optimization problem in variable-size design space. The goal of the study is to investigate various approaches to adapting the basic genetic algorithm (GA) to switch between design spaces of different size when solving a mixed optimization problem in a variable-size design space. Methods. This paper considers and explores adapted versions of GAs that enable solving mixed optimization problems in variable-size design spaces. These GAs involve introducing into the chromosome either a single additional gene encoding an integer dimensional variable, or several additional genes encoding binary dimensional variables, or an additional tag vector associated with the chromosome. Dimensional variables and tag vectors allow one to control switching between variable-size design spaces during the optimization process and determine which genes in the chromosome are active and which are passive. Moreover, all GAs implement the encoding of parameters of different types using Gray code. Results. The results of the experimental studies, obtained using the tuning information retrieval results reranking models as an example, confirm the feasibility of using adapted GA versions for solving mixed optimization problems in variable-size design spaces. The besttuned information search results reranking model is integrated into the RAG system in order to produce a more accurate relevancy assessment of the already found information. Conclusions. Adapted GA versions allow switching between variable-size design spaces during mixed optimization, ensuring simultaneous search for optimized parameter values in variable-size design spaces and obtaining a higher MAP metric value compared to the value of this metric in an untuned model. Using a tuned information retrieval results reranking model in the RAG system improves the quality of generative language model responses to user queries.