A simulation study of responsiveness-based virtual fencing: a reinforcement learning approach to precision livestock management

Virtual fencing removes the need for physical barriers in grazing systems by defining a GNSS-based boundary. Animals wear GNSS-enabled devices that emit an audio warning as they approach the boundary and escalate to a brief electrical stimulus if they continue into the exclusion zone. However, most systems apply one-size-fits-all cueing despite substantial individual differences in responsiveness, which can concentrate shocks in slower-learning animals and raise welfare concerns. Here, we developed a biologically informed simulation in which a reinforcement-learning virtual fence controller learns responsiveness class-aware cueing policies for high, moderate, and low-responsive cattle agents. Using Q-learning, the controller selects acoustic cue parameters and determines when escalation to electrical stimulation is necessary. Across 150 independent training runs, the highest-scoring learned Q-learning policy achieved effective control within the modeled environment with very low shock use: 0.095% of cattle-agent cue interactions overall (0.055% high-responsive, 0.090% moderate responsive, 0.139% low-responsive). In that illustrative run, the learned policy produced statistically significant differentiation in cue frequency (p = 0.003) and amplitude (p < 0.001) across simulated responsiveness classes. These simulation results suggest that reinforcement learning can identify responsiveness class-aware virtual fencing policies within a biologically informed modeled environment, producing differentiated acoustic cueing strategies while maintaining effective control with low simulated use of electrical stimulation. However, the findings should be interpreted as proof of concept rather than evidence of field-ready cue optimization, and field validation would be required before conclusions can be drawn about real-world effectiveness, transferability to cattle, or welfare outcomes.