Hierarchical Population Game Models of Machine Learning in Control Problems Under Conflict and Uncertainty

Abstract

AbstractProblem statement: Currently, the Machine Learning Control (MLC) methodology is rapidly developing, using machine learning methods to solve problems of complex dynamic systems control. One of the main paradigms of MLC is the technology of neuroevolutionary synthesis of control and decision-making models, which is considered as a promising means of on-board implementation of intelligent control algorithms for autonomous systems under environmental uncertainty. At the same time, when using the neuroevolutionary approach to solve multi-criteria control problems for dynamic systems under conflict, the following problems arise. First, the problem of training an artificial neural network (ANN) should be formalized in the form of a multi-criteria optimization problem under uncertainty (MOU), for the solution of which it is necessary to use approaches that generalize the well-known principle of guaranteed result of Hermeyer Yu., and take into account the conflict nature of the MOU problem. Secondly, when forming models of dynamic systems neuro-control, the training set has a complex structure and can include components that are themselves sets represented in a continuous form. The need to solve these problems determines the high computational complexity of the task of training the ANN and the need to develop a more efficient computing technology compatible with promising computing architectures, coevolutionary models and methods of distributed computing. Purpose of research: Development, software implementation based on distributed computing and research of the effectiveness of hierarchical population game models of coevolutionary algorithms for solving the ANN training problem, formalized as a MOU problem. Results: A methodics of ANN training based on hierarchical population game models of coevolutionary algorithms for solving the MOU problem using the principles of vector minimax and vector minimax risk is developed. Practical significance: The developed hierarchical population-based game models of coevolutionary training algorithms for ANN can be used in the design of SEMS neural control systems under conflict and environmental uncertainty. The problem of multi-criteria synthesis of an ANN, which is an integral part of a hierarchical neural network ensemble and implements algorithms for optimal robust neurostabilization of an unmanned aerial vehicle in a wide range of changes in environmental conditions, is solved.KeywordsMachine Learning Control (MLC)Multi-criteria optimization under uncertainty (MOU)Principle of vector minimaxPrinciple of vector minimax riskHierarchical population game modelUnmanned aerial vehicle (UAV)Neurostabilization systemHierarchical neural network ensembleCoevolutionary algorithms