Dr. Stefano Albrecht and collaborators at University of Birmingham and University of Texas at Austin receive a $1M grant from the US Office of Naval Research (ONR). The 3-year project will develop new algorithms for explainable reasoning, learning, and interaction for ad-hoc multi-agent collaboration. Research at Edinburgh will focus on algorithms for scalable and robust multi-agent teamwork in complex environments under dynamic team composition and resource constraints, using techniques of deep reinforcement learning, multi-agent learning, and graph neural networks.
"Imagine a team of autonomous agents collaborating with previously unknown agent and human teammates in a disaster rescue operation. Such environments are complex, chaotic, and hazardous, with time-sensitive contingencies, resource constraints, and potentially adversarial actors. The agents’ actions are non-deterministic, and they make decisions based on different descriptions of uncertainty and incomplete domain knowledge in the form of facts, relations, rules governing domain dynamics, and statements that hold in all but a few exceptional circumstances. Humans will not have the time and expertise to provide comprehensive domain knowledge or elaborate feedback. In addition, these agents are not all programmed by the same people, and they have heterogeneous sensing and actuation capabilities, communication protocols, and world models that are not known to each other. Also, team composition and the capabilities of teammates may change over time. As a result, team strategies cannot be developed centrally or a priori. Each agent must instead be prepared to tailor its sensing, reasoning, and behavior at different levels of abstraction to collaborate as well as possible with different types of teammates: those with which it can communicate and those with which it cannot; those that are more mobile and those that are less mobile; those with better sensing capabilities and those with worse capabilities. Also, to promote accountability, reduce development time, and collaborate effectively, agents will need the ability to describe their decisions, beliefs, and experiences, in a suitable format and level of abstraction. Towards realizing this futuristic scenario and achieving long-duration collaboration in complex domains under such ad hoc teamwork settings, we propose to design agents equipped with multi-level, reliable, and explainable reasoning, learning, and flexible teaming skills, which support robust collaboration with previously unknown teammates over extended periods of time."