Collaborative Research: RI: Medium: RUI: Automated Decision Making for Open Multiagent Systems

点击翻译按钮获取中文摘要

Various types of uncertainties complicate decision making in real-world contexts. In addition to imperfect sensing, there is added uncertainty in shared contexts due to the unknown actions of others and the dynamism brought about by these agents. Open systems are those real-world contexts whose composition changes over time due to either internal or external events. This research investigates how decision-makers (i.e., agents) may best act under uncertainty in open systems. Three forms of openness will be explored. The first is when the agents enter or leave the system over time. The second occurs when the tasks that must be completed by agents change over time. The third occurs when the agents’ capabilities change from learning new roles or skills. All three forms of openness, though prevalent in the real world and found in examples such as human organizations, disaster response, and smart transportation, have not been studied previously with respect to how they complicate decision making and their important role in enabling applications of artificial intelligence. Researchers from the Universities of Georgia and Nebraska-Lincoln, and from Oberlin College, will collaborate on this project. A new evaluation initiative leading into the creation of a competition involving use-inspired domains exhibiting various types of openness will be launched to spur broader interest. An innovative lesson module based on principles of creative thinking that brings the challenges of openness and how we may address them to undergraduate and graduate students will allow this project’s outcomes to be integrated into the classroom.The project takes the approach of investigating frameworks for modeling the various types of openness and realizing methods for acting optimally in the context of these frameworks. Specifically, the researchers will continue their investigations into scaling automated planning and reinforcement learning to open systems involving many agents with a novel focus on understanding the impact of task and frame openness. The ultimate goal is to combine representations of all three forms of openness and study whether this makes the decision-making problem fundamentally harder. Synergies between the planning and learning techniques under each type of openness will be identified and exploited. When combined with the advances of the past couple of decades in decision making under uncertainty due to sensor noise, these methods will represent a transformative step in translating principled planning and learning to the true complexities of real-world contexts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.