Markov network versus recurrent neural network in forming herd behavior based on sight and simple sound communication

摘要

Sound emission based on information received from the environment, including messages made by other individuals, enables communication between organisms of a given type (e.g., victims). Sound is the main form of communication for animals that they can incorporate into the decision-making process. In this paper, we describe conducted experiments to observe the role of sound communication in forming herd behavior. During the simulation, we investigated prey and predator organisms steered by a controller in the virtual world. We consider two types of agent controllers. The first one is developed using a Markov Network, the second one - a Recurrent Neural Network. The controller, based on information received in the form of environmental stimuli or states of own memory, makes decisions to change the position or, optionally, to make a sound that can then be picked up by nearby individuals. To find the parameters of the controllers, they are evolved by a genetic algorithm. In each generation, genotypes are decoded to the recurrent neural network or Markov Network, then some steps of simulations in a unique artificial environment, modeling the real world, are performed. On this basis, the evaluation of individuals is calculated. The main research element in this work was examining the impact of simple sound communication on forming herd behavior under the predator pressure. A comparison of controllers, i.e., Markov Network and Recurrent Neural Network, was the second goal of our research. (C) 2020 Published by Elsevier B.V.