COMPARISON OF REINFORCEMENT LEARNING AND MODEL PREDICTIVE CONTROL METHODS FOR UAV ROUTE PLANNING

Abstract

The article is devoted to an overview and comparison of two methods of UAV route planning: Reinforcement Learning and Model Predictive Control, which is a hot topic in the modern UAV navigation industry. The purpose of this paper is to determine the specifics of the use, limitations and other properties of the above methods. The article presents the principles of the above methods. Methodology. The paper compares the characteristics of Reinforcement Learning and Model Predictive Control methods for UAV route planning. The evaluation was carried out according to the following criteria: type of environment, the need for an accurate model of the environment, adaptability in dynamic environments, computational complexity, and data requirements. The scientific novelty is a comprehensive analysis of the efficiency of Reinforcement Learning and Model Predictive Control methods, in order to obtain the most popular comparative characteristics necessary for choosing an method. Conclusions. For the Reinforcement Learning method, the role of the agent, feedback, and the impact of the reward calculated using the Q-function on the model learning process are described. For the Model Predictive Control method, the features of using linear and nonlinear models for predicting future behaviour in the receding horizon strategy and the associated high computational complexity are outlined. The key advantages and limitations of the aforementioned methods are presented. The choice of a suitable route planning method depends not only on the specifics of the task, but also on the requirements for computing resources and the real-time response capabilities of the system under consideration. The importance of integrating different approaches to achieve optimal results in difficult conditions is emphasised. According to the results of the study, the characteristics of Reinforcement Learning and Model Predictive Control methods for building UAV routes are presented by the type of environment, the need for an accurate environment model, adaptability in dynamic environments, computational complexity, and data requirements. Further research in this area could focus on improving methods and algorithms that allow UAVs to adapt to changing environmental conditions in real time, combining Reinforcement Learning and Model Predictive Control methods with other route planning algorithms, such as A* or Ant Colony Optimisation, to take advantage of each.