SBIR-STTR Award

SIFT, along with Drs. Julie Adams and Ramesh Sagili of Oregon State University (OSU), propose to develop Resilient Emergent Properties for Autonomous Agent InteRactions (REPAIR). Building upon recent innovative work in swarm control and optimization, and applying new observations of specific defensive and adaptive characteristics of honeybee colonies, REPAIR will provide biologically inspired metaheuristic control algorithms for autonomous agents, along with a framework for ongoing development and evaluation. Recent honeybee research has identified two remarkable phenomena: altruism, where individual bees under duress remove themselves from the colony in order to protect the hive, and drifting, where a bee who has separated from its original colony is gradually adopted into a new one after a courting process. The primary objective of REPAIR will be to capture, in theoretical and algorithmic form, the unique behaviors of altruism and drifting, and to demonstrate the beneficial effects of these capabilities in simulated scenarios of interacting autonomous agents, both in the robotic and cyber domains. A Phase I pilot study at the OSU honeybee lab will observe the process of altruistic self-removal under duress, and the open-source Robotarium platform will be used for preliminary evaluation of new metaheuristic control policies in the REPAIR framework.

Most AI agent systems are extremely brittle: they use hand-crafted or laboriously-learned models that cannot handle situations that deviate widely from their expectations or training set. Real-world applications required more robust and resilient methods. In the original REPAIR Phase I and Phase II projects, we developed general metaheuristics for decentralized multi-agent systems that are capable of improving system resilience, making them more likely to complete their task when faced with threat that can disable agents. For example, we analyzed groups of agents that perform self-removal when faced with a threat of an infection, and we found that self-removal can significantly improve the resilience of the multi-agent system. This new Phase II project will build on the initial REPAIR results by extending the definition of resilience to include the ability of an agent to carry out its task in the presence of novelty, by detecting, characterizing, and accommodating the novelty. Here, novelty is defined as situations that violate implicit or explicit assumptions about agents, the environment, or their interactions. The DARPA SAIL-ON program has developed suitable evaluation domains and an evaluation process to quantitatively assess agent resilience to various kinds of novelty. SIFT has developed a new OpenMIND agent that introduces model revision functionality into the central deliberation loop, allowing the system to always "keep an open mind." SIFT's agent has been successful in SAIL-ON evaluations, meeting program metric targets for novelty detection and adaptation across three different application domains. This Phase II project will create a unique opportunity to extend and potentially transition our prior work on REPAIR and SAIL-ON. The initial OpenMIND agent will be extended, improved, and evaluated in multiple domains with new kinds of novelty it has never handled before. The effort will develop new novelty detection and accommodates techniques for novel changes in environments, goals, and events. We have also included an Option to support technology transition via the SAIL-ON Capstone Demonstration, in which our enhanced agent technology will be applied to a realistic DoD application area and evaluated against less-adaptive state-of-the-art solutions.