Genetic Optimization of Discrete High‐Level Policies for Swarm Shepherding with Angular State Abstraction

This paper investigates a compact and interpretable framework for swarm shepherding based on the genetic optimization of discrete high‐level policies in noisy flocking scenarios with stochastic sheep responses. Conventional rule‐based strategies, such as the Strömbom controller, provide effective and interpretable collect–drive behavior, but rely on manually designed rules and offer limited flexibility. To preserve interpretability while enabling data‐driven refinement, we represent the shepherd controller as a discrete state–action table defined by coarse geometric relations among the dog, the flock, and the goal, including the dog's angular sector relative to the flock–goal axis, flock dispersion, whether the dog is too close to the flock, and whether all sheep are already inside the goal region. Each state is assigned one of five high‐level actions: stop, clockwise orbit, counterclockwise orbit, collect, or drive. The policy table is optimized using a genetic algorithm initialized from a rule‐based prior derived from the Strömbom strategy and refined through elite preservation and mutation. Experiments compare the rule‐based controller with discrete‐policy variants differing in angular resolution and action‐set design. The results show that the proposed framework improves goal‐maintenance stability while maintaining competitive arrival performance, with a compact and interpretable policy representation.