Giving wildlife conservationists some helping PAWS

Ashley Kilgore:

Wildlife conservation is an enormous global undertaking, vital to ensuring the health and longevity of our planet and that the incredibly diverse plant and wildlife species we share our world with, are here for generations more to come. A significant threat to conservation efforts is the poaching of wildlife, which can be difficult and even dangerous to combat, and is pushing many species toward extinction, while also helping to support a multi-billion dollar illegal wildlife trade.

Ashley Kilgore:

I’m pleased to introduce my guest for today’s podcast, Lily Xu with Harvard University, whose work to create a data driven approach to combat poaching in protected areas around the world, led to the development of PAWS, the Protection Assistant for Wildlife Security. Perhaps most exciting is the hope that with the help of PAWS, not only can we reduce the impact of poachers, but ultimately reintroduce populations of wild tigers back into areas where they once thrived. I’m also thrilled to share that Lily’s work was awarded the Informs Doing Good with Good OR Prize, which recognizes outstanding student projects with a significant societal impact. And I’m excited for the opportunity to speak with her about it. Hi Lily, thanks so much for speaking with me. I can’t wait to dive into your work.

Lily Xu:

Thanks so much, Ashley. I’m honored to be here with you today on this podcast as a member of the Informs community. And I’m excited to share the work that we’ve been doing for the past three years.

Ashley Kilgore:

So now obviously conservation is a multifaceted complex effort with poaching being just one of the many challenges facing those trying to preserve and protect our natural world and its resources. Can you tell us, what part of the world in particular your work and that of PAWS has been focused on and why?

Lily Xu:

Thanks for that question, Ashley. So I’ll start off by saying that the work that we’ve done so far with PAWS, has focused on working with two parks, one in Uganda, and one in Southeast Asia, specifically in Cambodia. But part of our goal with PAWS is to build a system that is as generalizable as possible, so that it can be adapted to and used by parks all throughout the world, regardless of exactly their species and biodiversity makeup. So that’s the first thing I want to say, but then going back to the parks that we have been working with so far, there are parks in Uganda, which is home to a population of a lot elephants and giraffes. These parks are called Murchison Falls and Queen Elizabeth National Park in these rich African Savannah. And then this park in Cambodia that we’ve worked very closely with and deployed a number of few tests is Sre Pok Wildlife Sanctuary right on the border with Vietnam.

Lily Xu:

And it’s home to a population of elephants, deer, pigs, monkeys and birds and Sre Pok was actually a home to the last population of tigers in all of Cambodia. So the last tiger was spotted in Sre Pok back in 2007 but tigers are now believed to be extinct in the entire country. So unfortunately we’re already too late to save the native population of tigers there. But one really optimistic thing that personally drives me a lot in trying to push this work, is that WWF has identified Sre Pok Wildlife Sanctuary as the most promising site for tiger reintroduction in all of Southeast Asia. So we really want to ensure that our PAWS system is as effective as possible, to help these rangers prevent poaching, such that when the tigers are reintroduced there, they aren’t immediately poached away and that they also have a sufficient prey population to prey upon when they do arrive.

Ashley Kilgore:

That’s just so exciting that there is the hope of bringing back these populations of wild tigers. So can you share with us, what are some of the challenges associated with combating poaching that make this such a significant issue?

Lily Xu:

So one of the hardest things about poaching and trying to use an AI approach to combat poaching is that the AI systems, machine learning systems, optimization systems that we try to develop, require a lot of data. So we need to have a good understanding of exactly what the situation is like on the field, such that we can build a predictive model, we can have something to optimize. But the challenge is that poaching is such a hidden activity. So we aren’t able to observe most of it. And the poaching instances that we are able to observe, are extremely biased. So the only poaching that we are able to see, that we are able to track and keep note of, are the ones that rangers observe in conducting patrols around these parks, is when they are walking through one park and they see a wire snare that was laid on the ground.

Lily Xu:

They see a bullet cartridge, they see a poacher’s illegal campsite, and they’re able to make note of these observations. But you can imagine that if our data is suggesting that all the poaching is occurring in one specific region, because that’s what we’re seeing, that’s not necessarily reflective of the true situation on the ground. So it could be the case that there’s plentiful poaching activity going on elsewhere, that rangers are just not observing and perhaps they haven’t patrolled in that area very frequently. And that is something that we’ve noted a lot in all of the park feed that we’ve worked with so far, that the historical major patrols are extremely biased. That it’s much more concentrated in some regions. There’re many parts of the park that they’ve never patrolled in the past. So we have all of this uncertainty in the data and it just prevents us from being able to build a complete model.

Lily Xu:

So that challenge of the uncertainty of the data, imbalance and bias in historical patrol observations, was sort of the driving domain challenge that led us to one of our algorithmic approaches, that I think will discuss more a little bit later. But one other challenge that’s kind of unique to poaching and other kinds of observations, is that we have what some folks in computer science like to call positive and unlabeled data. So when you see a snare, we know for a fact that there was poaching activity there. But if the rangers conduct patrols through a region and don’t note anything, we don’t know for certain that there wasn’t actually illegal activity. It could be the case that there was illegal activity, we just weren’t able to see it. Perhaps it was behind a bush, perhaps you needed to patrol just a little bit farther or perhaps we needed to just time things a little bit differently. So whenever we have a negative instance of, there was no poaching, we can’t be 100% sure that it was a true negative instance, which makes this problem really challenging.

Ashley Kilgore:

So now PAWS represents a collaboration with a number of global organizations, such as the Worldwide Fund for Nature and the Wildlife Conservation Society. Could you share how this collaborative effort came together and what are all the moving pieces here?

Lily Xu:

Yeah, so the impetus of this collaboration first of all, our work on AI for poaching prevention and then subsequently our collaboration with these organizations, I have to give credit to my advisor, Professor [inaudible 00:09:27]. So he is a researcher who specializes in multi agent systems and game theory. And back in the 2000s, he had a long line of work with the US Coast Guard, trying to help them plan Coast Guard patrols at various US port cities and time it in a strategic game thermionic way such that the patrols are randomized, so that attackers wouldn’t be able to just sort of follow the strategy of the coast guard and then respond. According to that, there this optimal game, the calculations that they were implementing. So then after this project was deployed with the US coast guard, he got really excited and started thinking, okay, what are other domains in which there might be opportunities to apply game theory to real life problems.

Lily Xu:

And then this topic of poach and prevention and this potential strategic interaction between a ranger and a poacher came up as a viable opportunity. And I know that Millin is really fond of animals in particular tigers, which are plentiful, are historically plentiful and treasured in his own country of India. So that really pushed this desire to pursue coaching prevention as an AI application. And then from there just through very long iterative conversations with conservation organizations. So we first started working with the Uganda wildlife authority and specifically with parks in Uganda. And then from there after working closely with one park showing success, having that kind of strong partnership built up, then we were able to just kind of establish trust, establish a presence within the conservation community. And then from there, we were connected with folks at the world wildlife fund and the world wildlife fund is part of this partnership that is called the smart partnership, which is a consortium of nine different conservation organizations around the world, including the wildlife conservation society, North Carolina zoo Panera society of London, Frank zoo, and so on.

Lily Xu:

So basically just really investing on our relationships from one collaboration, paid off a lot of dividends in giving us access to this community, enabling us to foster trusting relationships with subsequent organizations, much more easily from there.

Ashley Kilgore:

And now the work of PAWS is supported by two unique algorithms’ lizard and mirror, which are integrated with the software system for protected area’s management known as Smart, which as you just mentioned is also the name of the partnership of conservation organizations around the world. Could you share a little background on the functionality of each of these algorithms and how they interact with each other?

Lily Xu:

Yes. Great question Ashley and I, before talking about all these algorithms listed in mirror, I want to start off by saying the base functionality of pause is a machine learning system that is doing supervised learning. So we have the historical patrol observations and we’re trying to make predictions on where we think poaching is most likely to occur based on the geography of the land. So we have information about density, rivers, roads, elevation, precipitation, temperature, all that kind of stuff. And we’re trying to make predictions on potential, poaching hotpots in the future in particular, in areas that we haven’t explored very much. And this is a system that we’ve tested that we’ve shown to perform well, both in terms of computational metrics, such as AUC and F1 score, but then also in field test on the ground. But you can imagine that as we discussed earlier, an effective predictive algorithm, predictive machine learning system, like this would require a lot of historical patrol data in order to make decent predictions.

Lily Xu:

So that inspired the first line of algorithm innovation, which was our list algorithm in which were focused on the challenge in under-resourced parks or data scarce parks that may not have a lot of historical patrol observations. So for these parks, you can imagine that they only started conducting patrols this year, or they’ve only been patrolling for one region of the park. And now they’re building a new patrol post in another location, but they have no sense of what the poaching patterns are like there. So for these kinds of parks, in which we don’t have a good sense of exactly what the poaching patterns are going to be like, we want to take an adaptive approach that will both go to places that we know to be existing poaching hotspots, where there are going to be a lot of snares, but then also proactively explore and patrol under explored areas so that we can make our predictive model better so that we can make better predictions a few months down the road.

Lily Xu:

So this mirrors, the exploration exploitation trade off, that’s common in the online learning Subo of machine learning. And our lizard algorithm offers a way to trade off this balance between exploration of unknown areas versus exploitation of known pushing hotspots in a way that tries to remove and locate as many stairs as possible to avoid these kinds of short term losses. Because if we take a long time to explore and then just keep saying, we’re going to improve our predictive model, we’re going to improve our predictive model so that it could be better five years down the road. We might have certain species that already extinct by that point. So we really have to make sure that we’re not wasting too much time exploring, which is not the case in a lot of other applications of online learning, such as recommendation systems like Netflix and Spotify in which they are able to wait a lot longer.

Lily Xu:

The stakes aren’t as high as, as the domain that we have that’s what our lizard algorithm focused on. And then from there, we consider a more complex setting in which instead of just trying to build a good predictive learning algorithm, we recognize that the more that we patrol in a certain area, the savvy poacher might recognize that their snares are all being taken away and will respond accordingly. So they might move to another area. So, and then this becomes a game interaction between the ranger who is trying to remove theirs and the poacher, who is adapting according to the rangers decisions. And now the ranger needs to make decisions, preemptively anticipating how the poacher is going to respond, which brings us to this more complex setting of sequential decision making, so that’s where our mirror algorithm comes in to help us learn sequential policies of how the rangers should be conducting patrols.

Lily Xu:

But the challenge that we have here is that typical algorithms for sequential decision making often are taking approaches such as reinforcement learning, which is machine learning approach to learn exactly these kinds of sequential policies, but reinforcement learning requires that we have a good simulator of the environment so that we have a good sense of what the reward system is going to be like in the environment. This is how the environment is going to change from one state to another state with a lot of precision.

Lily Xu:

But because our understanding of the environment of behavior is uncertain. And because these things also might change over time, we can’t have very precise values. So we can’t say the impact of us patrolling a one extra hour in this area of the park has exactly this 0.45 effect on poach behavior. But instead we might more reasonably estimate that parameter to within an uncertainty interval. So instead of saying this parameter is 0.45 we could say it’s somewhere between 0.4 and 0.5. We’re not sure exactly what, but that gives us an uncertainty interval over which we want to be robust. So our mirror algorithm enables us to do robust patrol planning that will learn a policy as optimal as possible to anything within the uncertainty interval.

Ashley Kilgore:

Now in your work, one thing in particular you mentioned is the importance of collaboration between researchers and practitioners and bridging the gap between research and practical application. How pivotal a role does this play in the work being conducted by pause?

Lily Xu:

So I think that the bridging the gap between research and practice is really what should drive us as an AI, as an operations research community. Because if we want to work on interesting technical problems, we need to make sure that they are interesting. And what is interesting should be defined in terms of what is actually useful and important for applications that we care about. And hopefully some of those applications are for things that benefit the collective good for ends that benefit those who are most marginalized in society, including people from all parts of the world, people from all kinds of backgrounds and also animals who oftentimes don’t have a voice in any of this.

Lily Xu:

So I think that ideally all research, especially this kind of work in applied computer science that I am working in, should always start by asking the question of what algorithms would be useful for this application problem that we care about. So the application that really drives me is environmental conservation. So we ask what algorithms will be useful here and why can’t they be applied immediately. And that will help us identify what that gap is that needs to be bridged. So for example, the two approaches I just spoke about with lizard and mirror as the two algorithms that we’ve developed in our projects, all of those have stemmed from saying, well, we have this system that works well on the ground but there might be a specific use case where it doesn’t work for an important real world setting. So we had a good predictive learning algorithm, but a lot of products don’t have a lot of data.

Lily Xu:

So how can we proactively acquire data in the most helpful way? And that inspired lizard algorithm really trying to address these challenges and constraints faced by new parks, by parks that were expanding the parks that were building up their ranger teams and so on. And then from there, we were trying to answer this question of, okay, how can we effectively use sequential planning algorithms in these kind of complex settings, such as patrol, pushing prevention through ranger conducted patrols. But the challenge there is sequential planning requires us to do a lot of iterations in some simulator to learn a policy. But the challenge is the real world is not a simulator that we can just keep experimenting on. So we have to distill the real world into some toy simulator but it is never going to be a perfect mirror of exactly what the situation is like on the ground.

Lily Xu:

So let us instead go simulator that will encapsulate instead in a range of different real world in instantiations. And then we want to say, how can we be robust to any one of those possible real world scenarios? So all of our work has really been driven by this challenge of bridging the gap between research and practice. And if it’s safe for me to say, I think that these have led to a really, really interesting technical problems and solutions in the work that we’ve done so far.

Lily Xu:

I think there’s a fear among a lot of researchers that it might be a waste of time to really delve in deep into these kind of domain challenges to really take the time to talk with domain experts and understand what is going on the ground but I think that it’s actually really valuable and the reward is that not only do we get to work on problems that matter more, but then also sometimes they will uncover a lot more interesting technical problems, more so than if we were just sitting in a block box, trying to just decide, okay, what new challenge do we work on arbitrarily.

Ashley Kilgore:

So Lily your work also mentions the critical role that rangers and park managers play in conservation management. How do pause algorithms amplify this valuable existing resource?

Lily Xu:

That’s a really great and important question. And I would say that the PAWS algorithm only exists because of the expertise of Rangers and conservation managers, and then the decisions and the predictions that we make are only intend to be an assistive aid, a decision making aid for the conservation managers on the ground. So we’re not trying to replace their expertise by any means but rather take their domain knowledge that has been integrated into the system. So the predictive albums that we’ve developed are learning has been designed in close collaboration and discussion with these park managers. So through conversations with them, we’ve grown to understand what is actually important. What are the things that drive poaching. For example, we learn that it’s necessary to build separate predictive models for the rainy season versus the dry season in a lot of parks, because for example, Cambodia experienced strong seasonality where certain regions, the rivers will completely flood over during the rainy season.

Lily Xu:

So those areas become impermeable at least by foot, but then during the dry season, they all dry up and then you can easily walk through it by foot. So it really changes the landscape in terms of what is accessible, where the animals are and so on. And therefore also impacts poaching behavior. So poaching patterns therefore vary a lot depending on whether it’s may season or dry season. So we learned that we had to build separate predictive models based on this domain insight that came from the conservation managers. But then after we generate predictions, after we recommend locations for Rangers to patrol our goal is to only act as a potential nudge to point them, to areas among various candidate sites that they might already be considering. But of course so much happens on the ground. For example, even if they’re trying to patrol one area that we recommended the Rangers might and should override are our recommendations based on the feedback that they’re getting on the ground.

Lily Xu:

So oftentimes they might find footsteps on the ground and should follow the footsteps. They might see that a Bush has been pushed aside and a motorbike trail is going through that area. Depending on how fresh that looks, maybe they should follow that if they hear gunshots in the distance or see an animal, they there’s just so much expertise that they have that cannot be replicated by any sort of automated computerized system. So really what we’re our objective is to just sort of augment their existing skills and help them make the most out of the data that they have from his patrol observations, and then point them to some insights on top of everything that they are already doing.

Ashley Kilgore:

And now, Lily, could you share, what has the impact of pause and its efforts been so far?

Lily Xu:

Thank you for asking that Ashley. So the impact of PAWS has a work along a few different scales. So the first being on the smaller scale, working with individual parks, we’ve conducted several field tests in both two parks in Uganda, Murchison falls and Queensland with national park and then also street park, wildlife sanctuary, and Cambodia in which we were going through this exact process of building predictive models based on their historical patrol data and recommending sites for them to patrol. And then in the field tests that we deployed in Cambodia, we kept these blind experiments. So we weren’t telling them, okay, this is what the machine learning algorithms predictions are being. We were just recommending various sites. And through following our recommendations in a single month, they were able to collect and remove over a thousand snares, which it compared to a normal month, they only find between 200, 500 snares.

Lily Xu:

So that was something that was really exciting to both us and them and furthermore pointing them to some areas that were under explored in the past and saying that this site seems to be a really potentially viable hotspot. Specifically. There was one in the Northeast region of Cambodia, where they had really not patrolled at all in the past five years of patrols. And then this was something that our PAWS system was predicting to be high risk. And we pointed to them to this location and they patrolled there and they reported back to us. We just found a ton of elephant traps there. And apparently what was happening is that a lot of pictures were coming in from across the international boundary with Vietnam and placing elephant traps there. And then coming back for the chaps, because oftentimes the big game hunters for things like elephants, rhinos and tigers are often an international POS who come from these much more larger structured poaching crime scene, the kids.

Lily Xu:

So that was something that they had been missing before that PAWS was able to help with, but then moving on to a larger scale. So the conservation impact of PAWS is something that we’re really trying to scale up as much as possible. So I mentioned before the smart partnership is a consortium of nine different conservation NGOs. And one of their main products, main projects is a software system that they’ve developed called also called Smart, which is a conservation management tool that enables conservation managers to record all the historical observations right now, smart can do all these features like query and visualize and help plan patrols, but there is no predictive or prescriptive component. They can sort of like draw things out. They can do searches, they can draw dots on the map but there’s no way to extrapolate. Based on patterns that we’ve seen before, what do we think poaching hotspots might be in the next month in these areas that we haven’t explored yet.

Lily Xu:

And how should we adapt our patrol plans accordingly. So we have integrated our PAWS system into Smart. It’s currently available in this beta version of smart that is accessible to parks right now with the goal that PAWS will eventually be available to all 1000 parks around the world that are using the smart system. And right now we are in the process of working with two parks, one in bellies and one in central America and the other in the, in KA national park in Zambia. So we’re working with these two parks to really test out our PAWS system in different settings. One it’s being driven by users who are making these predictions rather than on our side. So all of these are small steps to hopefully building up a system that is robust, flexible, and adaptable to a wide variety of geographies and landscapes and coaching behaviors. And hopefully becomes a useful asset in the large number of things that conservation areas are doing for protected area management.

Ashley Kilgore:

So Lily, as I mentioned in my opening and after hearing about your work, it should come as no surprise that it was recognized with the informs doing good with good or award at the informs annual meeting this past fall in Anaheim, California. So first of all, congratulations, and secondly, I’d love to hear how has having your work recognized in this way impacted you and your path forward?

Lily Xu:

Thank you so much for that. Congratulations, Ashley, that means a lot from that means a lot to me and I was very humbled surprised to have received this recognition, especially after seeing the presentations of other incredible finalists who were doing work on social choice theory on COVID testing. And I thought almost for certain that COVID has been the defining challenge of the past two years. Obviously this award is going to go to a COVID person like who is going to care about the animals in a place that they’ve never been to in the world in the midst of so such an urgent global pandemic.

Lily Xu:

That was humbling and exciting to just see this excitement for conservation as well and of so many other challenges. But I think the biggest thing for me in receiving this recognition is that first that it provides me a good deal of validation because in computer science in academia, it kind of feels that like so much of what I’m doing with pause is unrecognized because all of this work that we’re doing with field tests, all of this work that we’re doing with deployment on Smart, in conversations with these, with these parks conversations with, with park managers, we spent a week visiting Sri Aqua sanction in Ento, all of this feels like invisible labor in terms of what the computer science community values.

Lily Xu:

So to receive the recognition and to have an academic entity like Informs say this work is valuable. All these invisible things that typically aren’t recognized by academic reviewers is actually something that we want to validate and encourage more people to do. That is really exciting and I hope that this Informs doing good with good or award. It just becomes one of many other recognitions that people can get for kind of doing work that is beyond the typical technical algorithm development and code experiments that is common in computer science. And I hope that this is just an incentive and impetus for others to kind of take the time to pursue these socially impactful projects in conservation and in public health and whatever domain they find important.

Ashley Kilgore:

Lily, it’s truly been such a pleasure speaking with you. Thank you so much for sharing your award-winning work with our listeners before we wrap up our discussion. I’d love to hear what’s next for you.

Lily Xu:

So what’s next for me is still trying to push pause as much as I can. So I mentioned that we have all these ongoing work that we’re doing with integration with Smart, in doing more on the ground as in more parks, but on a more personal level, I am still continuing my PhD, I think for the next two years. But then around that time, I’ll be looking into academic positions. So trying to stay in the research world, trying to keep up this momentum in doing this kind of AI research for societally impactful application areas. So I really hope to just be working on conservation problems and other challenges with a technical spin for the next long while.

Ashley Kilgore:

Well, that’s super exciting. I wish you the best of luck and I’m sure you’ll see nothing but success in this next stage.

Lily Xu:

Thank you so much, Ashley. I really appreciate it.

Ashley Kilgore:

Want to learn more visit resounding the human.com for additional information on this week’s episode and guest, the podcast is also available for download or streaming from apple podcasts, Google play, Stitcher and Spotify. Wherever you listen, if you enjoy resounding the human, please be sure to leave review to help spread the word about the podcast until next time I’m Ashley Kilgore in this is resoundingly human.