Anti-Poaching as a Partially Observable Stochastic Game

International audience In today’s world, endangered species are threatened by widespread poaching, requiring intel-ligent land patrol strategies to effectively detect and prevent such activities. Several recentworks have developed game-theoretic models for anti-poaching, wherein determining equilib-rium strategies, often based on the Nash Equilibrium (NE) 1 , leads to effective patrol strate-gies. Additionally, due to the complexity and imperfect knowledge of the models, Multi-Agent Reinforcement Learning (MARL) methods are usually proposed to learn these strategies.Yet, even with anti-poaching emerging as a popular domain for MARL, the absence of botha general model and a publicly accessible implementation has hindered both the evaluationand development of new solutions. In this context, the objective of this work is two-fold: (i)formalize anti-poaching as a Partially Observable Stochastic Game (POSG) capable of gen-eralizing existing models; and (ii) provide a publicly available implementation of this POSG inPettingZoo (one of the most popular APIs to implement MARL environments).