Protecting the Lion’s Share: Mitigating Illegal Wildlife Poaching with Multi-Agent Reinforcement Learning

Abstract

Illegal wildlife poaching is the main threat to national parks across the world, destroying global biodiversity through species extinction and destabilizing vulnerable communities. To protect endangered wildlife from poaching, national parks employ rangers to conduct patrols, but there are often insufficient rangers to fully patrol parks that can span thousands of kilometers, leaving significant parts of the parks undefended. Previous work has developed data-driven policies for planning anti-poaching patrols that support park rangers in optimizing their limited resources, however these approaches do not account for the sequential nature of park patrolling. More recent research has begun to explore reinforcement learning, a method that enables sequential decision making, in the park domain, but these algorithms do not scale efficiently to the area of these national parks. In this thesis, we use recent advances in multi-agent reinforcement learning to create a scalable method for park patrol planning. We (i) define a Markov decision process for the park environment and a compatible problem setting for the park patrols operating in it. We then (ii) design and implement the first multi-agent reinforcement learning algorithm for developing park patrolling strategies for independent patrol posts. Finally, we (iii) experimentally compare the performance of the multi-agent algorithm against an analogous single-agent across different park sizes and patrol post configurations to show that multi-agent methods can approach the single-agent returns with far fewer compute resources. By introducing multi-agent reinforcement learning into the anti-poaching domain, we can leverage the existing operational divisions in national parks to create computationally feasible patrolling policies. This work paves the way for future research into distributing patrol planning policies in a scalable way towards the protection of endangered wildlife.