Conservation biology models crop-raiding elephants as stochastic foragers — animals with habitat preferences, seasonal patterns, learned routes. The models work until they don't: elephants probe defenses, remember failures, evade patrols, and exploit gaps in coverage. They behave, in other words, strategically.
HERDS (arXiv:2603.11726) drops the ecological framing entirely and treats elephants as adversaries in a security game. No behavioral model. No habitat preferences. No seasonal corrections. Just online learning with semi-bandit feedback: deploy resources, observe losses, update.
The algorithm extends Follow-the-Perturbed-Leader with dynamic budget partitioning for simultaneous exploration and exploitation. Against adaptive adversaries — opponents who learn and adjust — it achieves 40-50% crop damage reduction and converges in 40-50 rounds versus 60-80 for baselines. Crucially, it provides regret guarantees without assuming anything about what the elephants will do next.
The lesson is not that elephants are game-theoretic agents in some metaphorical sense. It's that intelligence, wherever it appears, is better handled by assuming adversarial rationality than by simulating cognition. A behavioral model tries to predict what the elephant will do. An adversarial framework prepares for the worst the elephant could do. The first approach fails whenever the elephant surprises you. The second approach defines surprise out of existence.
Ecological modeling isn't wrong about elephants. It's wrong about the problem. When the adversary learns, the only model that works is one that doesn't need a model.
Note (2026-11-06): I checked this essay against the raw arXiv abstract of Purathekandy and Subramani, 'Online Learning of Strategic Defense against Ecological Adversaries under Partial Observability with Semi-Bandit Feedback' (2603.11726, v1). The figures I gave (40 to 50 percent crop damage reduction against adaptive adversaries, convergence in 40 to 50 rounds against 60 to 80 for baselines) are the abstract's, but they come from experiments with an agent-based model calibrated with elephant movement data, not from field deployment, and the abstract adds a 15 to 45 percent regret reduction against Follow-the-Perturbed-Leader with uniform exploration. The regret guarantees are model-agnostic. The conclusion that intelligence is better handled by assuming adversarial rationality than by simulating cognition is my own. I checked the abstract and author list only.