Explainable and Behavior-Aware Decision Making in Self-Organizing Swarms

Self-organizing swarm systems of autonomous agents (peers) achieve their missions collectively relying on solely internal interactions without centralized coordination. The inherent nature of swarms lies in their decentralized, collective behavior supported by ad-hoc architecture, enabling agents to dynamically join or leave the swarm without any impact on its overall mission performance. This collectively demonstrating intelligence and problem-solving abilities surpass those of individual agents. However, self-organizing nature of swarms introduces challenges in ensuring consistency and explainability of run-time decision-making processes due to probable “black box” nature of the embedded AI/ML models. Peers in the swarm are intended to make adaptive decisions guided by distributed intelligence to handle events while maintaining interpretable reasoning pathways for reinforcement learning upon completion of the mission. Explainable event handling assumes providing transparent insights into the swarm’s emergent behavior and decision-making strategies, which can be complex and unpredictable, and cannot be inferred from the properties of individual peers. The paper introduces a framework that ensures explainability of the swarm behavior by introducing a formal language for specifying event signatures, adaptive probabilistic methods for run-time event detection, and an event log manager that incorporates new event categories while continuously updating the event probability space.