Overview of modern reinforcement learning methods for solving problems of optimal control of dynamical systems

Context and relevance. The tasks of synthesizing optimal control for nonlinear dynamic systems with constraints and uncertainties remain computationally challenging, especially in aerospace applications. Reinforcement learning is considered a practical tool for building feedback and/or accelerating planning when classical methods are difficult to apply. Objective. To systematize classes of algorithms for optimal control tasks and identify criteria for selecting an approach for a specific problem. Hypothesis. Practical applicability is ensured by correct formulation and consideration of requirements for control continuity, data, safety, and robustness; combined solutions are most effective. Methods and materials. A review and comparative analysis of families of different reinforcement learning algorithms was performed. Results. Actor-critic remains the basis for continuous control, while alternatives increase selective efficiency but are sensitive to model and data coverage errors. Conclusions. The most promising are hybrid architectures that combine reinforcement learning with basic controllers and ensure controlled compliance with constraints. The choice of method should be determined not only by quality, but also by safety, robustness, and computational cost.