iWare-E: An Update on the Adversarial Fight Against Poaching

This is a guest post by Cassidy McDonnell. See Cassidy’s bio below.

Unfortunately, the global poaching crisis described in Zimei Bian’s post has yet to be conquered and wildlife crimes are still prevalent in countries such as Uganda, Tanzania, Kenya, Zimbabwe, and South Africa. Species such as the oryx as while as black and white rhinos have already been poached to extinction. Not only do wild animals play a vital role in local ecosystems, but protected areas attract a significant number of tourists, who are important contributors to local economies, workforces, and GDPs. According to a report released by the International Institute for Environment and Development (IIED), on Wildlife Crime in Uganda, major poaching incidents have been recorded in 17 of the 23 protected areas in Uganda.

Park ranger resources are quite constrained in these protected areas, making it difficult to stop poachers from partaking in illegal activities in these places. Zimei’s post described the innovative modeling techniques developed by CompSustNet Associate Director Milind Tambe at the University of Southern California to optimize scheduling for these rangers and maximize the amount of protected area covered by safety officials.

Although PAWS and other programs like it have done important work to increase wildlife protection in Queen Elizabeth National Park in Uganda, Dr. Tambe along with Dr. Shahrzad Gholami and a team of collaborators have continued to address many of its issues with a new innovation: imperfect-observation aWare Ensemble (iWare-E).

iWare-E addresses and improves many of the inconsistencies and problems earlier models had not taken into consideration. For example, PAWS relies on a specific type of explicit attacker behavior for its model. Although this model can handle complex data sets, it is unable to do so on a large enough scale to be effective in national park settings. No program has yet been able to scale up to real world scenarios and each iteration of progress has been limited to scheduling for a single protected area.

Although larger-scale modeling has been attempted, these methods either result in solutions that are of low quality or require significant amounts of time and computing power to run, which are impractical for low resource outposts where these technologies would most need to be implemented. Finally, PAWS and other state-of-the-art models record data on an annual basis, which fails to account for short term poaching patterns.

In contrast, iWare-E evaluates poaching activity seasonally with a timestep (an interval for measurement) that lasts for three months as opposed to a year like in previous models. This model is also less computationally expensive because it includes a scalable planning algorithm that applies a piecewise linear approximation, a change that has resulted in a 150% improvement in solution quality and a 90% improvement in both accuracy and run time.

To use iWare-E, the protected area is broken into a grid of squares that each have an area of one km2. Each square has distinct values describing its terrain, distance values, animal density, and patrol effort (measured by the amount of distance traveled by park rangers across a cell during a specific timestep). An initial training dataset is input into a training algorithm in order to build a matrix to model the collected data, taking into account that not all signs of illegal activity (i.e. snares and traps) will be discovered by the park rangers due to the limited personpower and the hidden nature of many of the traps. This algorithm outputs classifiers and a binary vote qualification matrix that are input into a second algorithm to predict the probability of crime observation in a certain area. More information about these models and results can be found here.

iWare-E is revolutionary in its ability to efficiently account for imperfect crime information and uncertainty while modeling complex data. It has been able to produce effective patrol routes in a way that has been less computationally expensive than other similar state-of-the-art models as well as improve accuracy and runtime. As a result, iWare-E has become the first adversary behavioral model for wildlife protection to be run in multiple locations, having been tested in two different protected areas within Uganda. To continue work in the future, the Teamcore group may look to transfer knowledge between areas with rich data sets (i.e. Uganda) and areas where data is much more scarce (i.e. Cambodia). This growth will help expand research and improve solutions across different domains.

Cassidy McDonnell is a May 2019 graduate of Vanderbilt University in Civil Engineering. She became interested in Computational Sustainability through Vanderbilt’s University Course on the Ethics of Artificial Intelligence, which included a module on Computational Sustainability. Cassidy can be reached at cassidy.a.mcdonnell@vanderbilt.edu. The opinions expressed herein are Cassidy’s and do not necessarily represent the opinions of Cornell University.