Classification of Multi-Agent Systems by Moving Trajectory Using Neural Network Regression Model

Abstract

In this paper, we aim to classify multi-agent systems which consist of moving bodies, such as biological swarms and robot swarms, based on their trajectory data. Generally, the number of agents differs from system to system and the dimension of trajectory data is not constant, making it difficult to use the data as a direct input for a neural network classifier. In addition, the index in the data is unrelated to the connection between agents, and thus it is difficult to reduce the dimension of data using a simple algorithm. To solve these problems, we develop a regression model whose dimension of parameters is constant despite number of agents, and use the parameters of the model as an input data of a neural network classifier. In order to make the dimension of parameters constant, the proposed regression model uses the same parameter to calculate the interaction with another agent. The performance is shown through numerical experiments with observed data of biological swarms and computer simulation data.