Multi-agent control and intelligent sensor allocation with Genetic Programming and Reinforcement Learning

"The US Navy has identified the need to develop quantitative frameworks to solve a variety of challenging problems, including sensor management and allocation, and adaptive flight control of Unmanned Air Vehicle (UAV) swarms. In collaboration with the MITAI Lab, we propose to develop a framework that combines Reinforcement Learning (RL) and Genetic Programming (GP) to evolve adaptive solutions to these problems. The RL+GP framework will scale up to realistic problems, and will be analytically tractable.The proposed framework will also make it possible to conveniently incorporate domain-specific information, giving it the capability to develop "customized" solutions to each problem at hand. We will also develop an Agent-Based simulator of swarms of UAVs,and use the simulator as a testbed for the RL+GP framework. The simulator will allow testing of various sensor configurations, specification of UAV missions, and will include the ability to simulate a variety of normal and adverse scenarios. The simulatorwill also serve as a tool for visualization and for quantitative evaluation of the proposed framework or other flight control and sensory allocation technologies. The RL+GP framework combines the power of learning and of evolutionary computation, but isanalytically tractable and simplifies the process of encoding domain information, more so than other adaptive frameworks. Coupled with the Agent-Based UAV simulator, we expect this framework to be applicable to a wide variety of problems of military andcommercial interest."