On the Use of Echo State Networks in Various Configurations to Predict the Dynamics of Adversarial Swarms
Adversarial (competitive) swarms consist of two or more systems (each system consisting of a collection of individuals, interconnected agents) where the goals of each group are conflicting. This work aims to use an Echo State Network to predict the individual behavior of agents in two adversarial swarms and thereby develop an improved understanding of the dynamics of such systems. The current study was divided into three phases. An agent-based Adversarial swarm model was initially developed comprising of two competing swarms, the Attackers, and the Defenders, respectively. The Defender aimed to protect a point of interest in unbounded 2D Euclidean space called the Goal. In contrast, the Attacker’s main task was to intercept the Goal while continually trying to evade the Defenders, which get attracted to it when they are in a certain vicinity of the Goal. The simulation was considered Semi-Hybrid as agent compromise, and goal compromise criteria were modeled to introduce realism as real-world engineering applications. The final system state was studied for all the varied number of agents making up each swarm. The effectiveness of the Semi-Hybrid approach was validated by using Multiscale Entropy, which revealed a greater degree of randomness for the Defenders than Attackers. In the second investigation, two configurations were used to evaluate the use of Echo State Networks for predicting group dynamics for each swarm. Configuration 1 employed a single ESN, i.e., the patio-temporal data for all agents of an Adversarial Swarm model was used input. In configuration 2, two separate ESNs, in parallel, were used to predict Defender and Attacker swarm dynamics. It was concluded that the parallel ESN configuration was more effective in achieving qualitatively similar predictions of the dynamics for the Adversarial Swarms. In the final investigation, an instance of an ESN in a massively parallel framework was trained on individual spatio-temporal data of every agent. The optimal hyperparameters obtained for every individual agent in the framework showed considerable variance that implied every agent in the Adversarial swarm reacted uniquely when a uniform stimulus was applied and thus reaffirmed the concept of individuality of agents in a swarm.