Enhanced Neural Network Learning In Imagination of Systems for Controlling Unmanned Movements of Objects

Abstract

Introduction. Artificial intelligence based on recurrent (cyclic) neural networks (RNNAI) performs the function of learning to think like the human brain. Large recurrent neural networks can learn spatial and temporal representations of data. The model-based reinforcement deep learning method requires high-dimensional input data to achieve an optimal agent action strategy and an accurate representation of the environment.

However, reinforcement learning faces the traditional problem of large data: algorithms rarely work with high-dimensional data. First, traditional RL algorithms have difficulty learning the weights of a large model in the task of assigning grades or credits, especially at the end of a sequence of algorithmic steps. The weight distribution problem solves the problem of determining which steps should be considered rewarding or punishing for the final result. Second, the performance of physical simulators, whose task is to anticipate changes in the environment, is low. The selection of optimal actions is carried out through separate task scheduling. This requires simulating many random actions and selecting the best one. This is classic Model-Based RL. However, with large dimensions and long chains, the number of possible actions becomes too large to enumerate. Therefore, the author explored an approach that uses a mental model of the world as the environment model. Neural networks in this approach are trained similarly to how they are trained in the human brain. This approach achieves the control goal by constructing a model of the world instead of conducting costly real-world testing.

Purpose of this article is to analyze current international experience in developing and implementing analytical platforms for controlling moving objects in single- and multi-agent systems. This control is achieved using artificial intelligence generated by deep neural networks, training the model through reinforcement learning in unknown, partially observable environments.

Methods. An approach to controlling moving objects in single- and multi-agent systems using neural networks and a mental model of the world is considered.

Results. An approach to controlling moving objects using neural networks and a mental model of the world is investigated. This article analyzes international experience in the development and application of artificial intelligence tools, specifically deep reinforcement learning, to solve problems of moving object behavior in unknown, partially observable environments.

Conclusions. Based on this analysis, the author proposes the application of a well-known approach based on deep reinforcement learning to the problem of controlling moving objects. This approach achieves the control goal by constructing a model representation of the world instead of conducting costly real-world testing.