Neural networks in agent-based models: advantages and disadvantages of hybrid research methods

The use of the agent-based approach to modeling socio-economic processes has recently become more frequent. The model close to reality allows us to assess the situation, simulate different scenarios of model experiments, and on the basis of the results to form recommendations to the authorities. The paper brings up the problem of describing agent-based patterns of behavior of economic agents with limited rationality, which necessitated the use of additional modeling techniques, namely neural networks. The purpose of this article is to analyze current experiences and prospects for the use of neural networks in agent-based modeling of socio-economic processes. The study considers examples of Russian and foreign hybrid agent-based models developed in the framework of public research at universities and research organizations and presented in articles and monographs available for public use. On the basis of generalisation of experience of domestic and foreign scholars on the use of neural networks in agent-based models the advantages and disadvantages of hybrid methods of researching socio-economic processes were identified, which reflects the novelty of the research. The paper describes the most popular current applications of neural networks in agent-based modeling of socio-economic processes. It covers the main ways of integrating neural network and agent-based models used in the current stage of technology development. The paper describes the local (micro-simulation) and global (macro-simulation) approaches to the use of neural networks in agent-based models and the methods of their implementation on the example of integrating machine learning in an agent-based model. Advantages and disadvantages of hybrid agent-based models of socio-economic development of territories are revealed, and they are reflected in the summary matrix of the SWOT analysis. The authors provide examples of specialized software that helps embed neural networks in agent-based models. Two options for embedding are described: online (using the same tool of a wider profile) and offline (using several software products). The prospects of using neural network technologies in agent-based models are due to substantial improving the quality of modeling and providing a more realistic behavior of agents.