Bayesian optimization of distributed neurodynamical controller models for spatial navigation

Abstract

Dynamical systems models for controlling multi-agent swarms have demonstrated potential advances toward resilient and decentralized spatial navigation algorithms. The bottom-up mechanisms of spatial self-organization in such models can produce complex or chaotic behaviors. For example, we previously introduced the NeuroSwarms controller, in which agent-based interactions were modeled by analogy to neuronal network interactions, including spatial attractor dynamics and temporal phase-synchronization, that have been theorized to operate within the hippocampal place-cell circuits of navigating rodents. While this complexity may drive emergent navigational capabilities, it also precludes linear analyses of stability, controllability, and performance that are used to study conventional swarming models. Further, tuning dynamical controllers by hand or grid search is often inadequate due to the potential complexity of objectives, dimensionality of model dynamics and parameters, and computational costs of continuous-time simulation-based sampling. Here, we present a generalizable framework for tuning dynamical controller models of autonomous multi-agent systems based on Bayesian Optimization (BayesOpt). Our approach utilizes a task-dependent objective