Bio-inspired, AI-based Systems for Intelligent Counter Swarm- BASICS

Abstract

Swarming unmanned aerial systems (UAS) are quickly becoming a preferredtactic for adversarial powers, with attacks on key U.S. allies showcasing theirpotential against soft targets. Improved endurance leads to UAS flying low andslow, arriving from unique routes, complicating ground based detection. Currently,swarm operators employ rudimentary navigation tactics which rely on a numbersadvantage, or repeated waves of assaults. However, swarm tactics using advancedalgorithms, like machine learning, are emerging, and defense methods will bemore routinely asked to operate against complex strategies.Many existing counter swarm algorithms suffer from a glaring problem, in whichthe number of intruders, their positions and trajectories, their goal(s), and theirmaneuverability are assumed to be known. Looking to nature for inspiration,a bio-inspired algorithm is developed, mimicking the hunting patterns of theHarris Hawk. This American raptor employs collaborative hunting strategies tomaximize prey exploitation in the scarce desert environment. The presented bioinspired, counter swarm algorithm uniquely captures the interplay between eachagent’s autonomy and multi-agent collaborative task allocation, maximizing theeffectiveness and providing the flexibility required for such a complex optimizationproblem. This work uses a two-part strategy: (1) Initial search: where a globalheat map is developed in real-time, with each agent contributing search knowledge.The heat map allows the integration of memory structures on a global scale andalso discourages searching recently visited areas. (2) Intruder information sharingand collaborative navigation strategies: which are used to influence collectivedecision-making towards successful agents to avoid falling into a local minimum.Within the dynamically changing environment of counter swarm, a fixed strategyof deterrence is costly and inefficient, and could be subverted by an intelligentintruder. The heat map provides a nonlinear and flexible approach to searching. Itprevents the premature re-visitation of previously explored areas and can also beused to return to suitable navigation paths to localize and intensify targeted search.Intruders identified through this search are then exploited through collaborativetactics, which look to leverage shared information between agents and their successin "finding and killing" to produce more favorable attack strategies.Due to the highly dimensionalized nature of this environment, and the complexityof balancing exploration and the development of search tactics, which need to findand kill an unknown number of intruders using multiple agents, a reinforcementlearning-based approach is uniquely adopted. The resultant algorithm is validatedusing a large number of randomly selected scenarios to assess its ability togeneralize the find-and-kill policy (a.k.a. tactics). Tests are conducted in largerareas, with varying numbers of agents and different velocity ratios, while alsoconsidering unique behaviors like a central swarm goal or split swarm. Otherwell-known methods (e.g., grid search, etc.) are used as the base for quantificationto assess the developed algorithms. The results demonstrate the effectiveness ofreinforcement learning-based algorithm to find and kill intruders compared toother commonly used algorithms