Самоорганизующиеся нейронные клеточные автоматы для обучения с подкреплением и эволюционного развития

Abstract

The relevance of studying self-organizing systems through the modeling of cellular automata is shown, which allows solving various problems from optimization and resource management to predicting the behavior of systems. A review of cellular automata is provided, including classical cellular automata, variational neural cellular automata, and growing developmental networks with neural cellular automata. The significance of using cellular automata in modeling complex processes that require the ability of the model to recover and retrain when the properties of an agent or environment change is discussed. In particular, it is shown that the introduction of self-organization into the model of reinforcement learning agents endows them with adaptability properties, which makes it possible to make changes to the original problem without additional retraining of the model and restore the system’s performance in case of damage to its structure. Examples of the use of cellular automata for modeling various processes and training agents are given.