Decentralized Center of Mass Prediction Using Reinforcement Learning and Global State Prediction
Abstract: Reinforcement Learning has seen successful ventures in single robot systems with its ability to solve complex problems with robust solutions. The application of similar techniques to multi agent systems, specifically in decentralized environments, introduces greater complexity and non-stationarity which effects the efficacy in training and convergence to a solution. This non-stationarity can be solved through the introduction of some level of shared or global knowledge, but this introduction still can limit the learned model's ability to scale. To solve this issue of non-stationarity while keeping the robust scalability of decentralized swarm networks, I present using Neighborhood Global State Prediction (GSP) and local policy generation to train a decentralized robot network to come to consensus for a value prediction. I simulate an unknown center of mass object which requires each robot's active input to move the object, and use the KheperaIV GAARA gripping module to provide all necessary vectored force feedback to train. Utilizing Neighborhood GSP allows the robots to converge to a predicted center of mass, overcoming the noise and non-stationarity additional robots introduce.
Advisor: Professor Carlo Pinciroli
Committee: Professor Constantinos Chamzas, Professor Kevin Leahy, Professor Carlo Pinciroli