Project Type
group
Authors' Class Standing
Graduate Student
Lead Presenter's Name
Shane Stebler
Faculty Mentor Name
William MacKunis
Abstract
Methodologies aimed at Advanced Control through Learning in Autonomous Swarm Systems (A-CLASS) utilize the tools of nonlinear control to develop and rigorously analyze new strategies of machine learning for formation control in collaborative autonomous multi-agent UAV groups. In A-CLASS, new algorithms for online, real-time machine learning can be achieved through the investigation of new automatic control policy improvement strategies, which optimize future control actions based on past experience. Recent advances in reinforcement learning methods have made significant progress in understanding and mimicking brain functionality at the level of the brainstem, basal ganglia, cerebellum, and cerebral cortex. In particular, neural network (NN)-based actor-critic structures, which utilize approximate dynamic programming (ADP), are promising techniques for simulating brain-like thinking in engineering systems. In this research project, new experience replay-based multi-agent control techniques will be investigated, which ‘learn’ from experience in real time the optimal control action in response to a given sensor stimulus. The end goal of this research project is experimental demonstration of new A-CLASS-based control methods on multi-agent groups of autonomous quadcopter UAVs using a motion capture arena.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
Yes, Spark Grant
The A-CLASS Method for Autonomous Quadcopter Formation Control
Methodologies aimed at Advanced Control through Learning in Autonomous Swarm Systems (A-CLASS) utilize the tools of nonlinear control to develop and rigorously analyze new strategies of machine learning for formation control in collaborative autonomous multi-agent UAV groups. In A-CLASS, new algorithms for online, real-time machine learning can be achieved through the investigation of new automatic control policy improvement strategies, which optimize future control actions based on past experience. Recent advances in reinforcement learning methods have made significant progress in understanding and mimicking brain functionality at the level of the brainstem, basal ganglia, cerebellum, and cerebral cortex. In particular, neural network (NN)-based actor-critic structures, which utilize approximate dynamic programming (ADP), are promising techniques for simulating brain-like thinking in engineering systems. In this research project, new experience replay-based multi-agent control techniques will be investigated, which ‘learn’ from experience in real time the optimal control action in response to a given sensor stimulus. The end goal of this research project is experimental demonstration of new A-CLASS-based control methods on multi-agent groups of autonomous quadcopter UAVs using a motion capture arena.