Biomimetic Optimal Tracking Control using Mean Field Games and Spiking Neural Networks

This paper investigates decentralized optimal tracking control for multi-agent systems (MAS)s with a large population. Unlike conventional decentralized control, two major challenges must be addressed when the population size of the MAS is large: the "curse of dimensionality" and environmental uncertainties. The paper develops a novel online learning decentralized adaptive optimal control strategy to address these challenges by combining the emerging Mean Field Games (IVIFG) theory with a novel Biomimetic Actor-Critic-Mass (B-ACM) learning algorithm. Mean-field control is developed as a decentralized optimal controller that can effectively reduce the computational complexity and the communication effort. A Biomimetic neural network that mimics the human brain, which is much more efficient than traditional Artificial Neural Networks (ANNs), is designed using Spiking Neural Networks (SNN)s. The information is encoded into a sparse spikes vector similar to the human brain. The SNN technique and mean-field control are merged into one unified framework, B-ACM. The B-ACM includes three regions of neurons in coordination with mean-field control: 1) Reward region to approximate the optimal cost function, 2) MAS Population Estimation region to predict the effects from other agents, and 3) Action region to compute the optimal control. Moreover, the paper introduces a novel SNN weight update law based on gradient descent. The effectiveness of the proposed scheme is validated through numerical simulations. Copyright (C) 2020 The Authors.