Learning and Planning Under Uncertainty for Green Security

Lily XuIJCAI 2021 · OpenReview

Abstract

Lily Xu’s research addresses "Green Security"—the protection of wildlife and forests—by developing AI techniques that bridge the gap between theoretical game theory and real-world conservation. By integrating uncertainty-aware machine learning with robust sequential planning, her work enables rangers to more effectively monitor massive, resource-constrained landscapes. Field tests in Cambodia's Srepok Wildlife Sanctuary demonstrated a five-fold increase in snare detection compared to traditional methods, showcasing the potential for AI to transform global conservation efforts.

Introduction

Conservation efforts are fundamentally a game of cat-and-mouse played over massive, difficult-to-monitor landscapes. In the Srepok Wildlife Sanctuary in Cambodia, for example, just 72 rangers are tasked with patrolling 3,750 km². Poachers hold a significant advantage: wire snares are inexpensive and can be deployed in "blankets," while ranger resources are severely limited.

Rangers patrolling the vast and difficult-to-monitor terrain of a wildlife sanctuary.

Existing AI solutions often fail in this domain because they struggle with three core challenges: historical patrol data is often biased and sparse; poachers are adversarial and adapt their behavior to ranger presence; and every patrol must balance catching poachers today (exploitation) with gathering data for tomorrow (exploration).

Methodological Framework

Xu’s work spans the entire pipeline from data analysis to real-world deployment, utilizing several distinct algorithmic approaches to handle varying levels of data availability and uncertainty:

  • Uncertainty-Aware Ensembles: In data-rich environments, Xu uses Gaussian processes within an ensemble method designed to handle extreme class imbalance (where "no snare found" is the overwhelming majority). These predictions feed into a Mixed Integer Linear Program (MILP) to optimize patrol routes.
  • LIZARD (Lipschitz-continuous Arm-Decomposable Bandits): For data-scarce regions, the problem is treated as a multi-armed bandit (MAB) task. LIZARD allows for "dual-mandate" patrols that simultaneously detect illegal activity and collect valuable data. It leverages the assumption that rewards are "smooth" (Lipschitz-continuous) across space to generalize findings from one area to adjacent ones.
  • MIRROR (Minimax Regret Reinforcement Learning): To handle unknown poacher preferences, this algorithm formulates patrol planning as a zero-sum game against "nature." It uses reinforcement learning to find policies that minimize "regret"—the difference between the chosen strategy and the best possible strategy in hindsight—ensuring the plan is robust even if the model of the attacker is slightly off.

Key Results and Real-World Impact

The research relies on a mix of historical records from the Srepok Wildlife Sanctuary and integration with the SMART (Spatial Monitoring and Reporting Tool) platform, which is used in over 800 protected areas globally.

During a one-month field test in Cambodia, the AI-generated patrol routes led rangers to detect and remove over 1,000 snares. This was five times higher than the average monthly snare recovery rate for that sanctuary. Beyond immediate field success, the LIZARD algorithm achieved theoretical "no-regret" status, outperforming standard bandit approaches in high-stakes environments.

Rangers displaying a large collection of confiscated wire snares in Cambodia.

Discussion and Future Directions

A central insight of this work is the "Dual-Mandate" strategy: moving away from seeing patrols as purely defensive and instead treating them as data-gathering missions. This solves the "cold start" problem in new parks. Furthermore, the research provides technical evidence of a measurable deterrence effect, confirming that poachers are strategic actors who respond to increased ranger presence.

While highly effective, future work must address heterogeneous resources (accounting for different ranger skill levels and equipment) and execution uncertainty (re-planning on the fly when environmental factors like floods block optimal paths). Xu also emphasizes the need for a participatory approach to ensure AI deployment aligns with environmental justice and does not inadvertently harm local marginalized communities.

Summary updated

BibTeX
@inproceedings{xu2021learning,
  author = {Lily Xu},
  title = {{Learning and Planning Under Uncertainty for Green Security}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=hjOK2h5PCVb},
}
Figure