As part of our AI in Your Community series, I sat down with Fei Fang, an Assistant Professor in the Institute for Software Research in the School of Computer Science at Carnegie Mellon University. She works on game theory and machine learning, researching the strategic behavior of multiple agents, which has applications to many societal challenges including poaching, and infrastructure security.
Tara Chklovski: Tell me a little bit about what problem you are trying to solve.
Fei Fang: My research area is game theory and machine learning. I study how we can use AI to address societal challenges such as environmental sustainability. For example, protecting wildlife from poachers, protecting forests from illegal logging, and protecting fisheries from illegal fishing. I also work on infrastructure security. So, for example, how can we protect ferry lines from potential attacks? And more recently I’m also working on how to address challenges in transportation systems like making the ride-sharing platforms more efficient.
Although they look quite different, there is something common underlying all these problems – the strategic behavior of multiple agents.
For example, protecting wildlife from poachers is similar to protecting, say, ferry lines from potential attackers in the sense that we have a defender side, or law enforcement side, which needs to schedule their defensive resources (like the patrol boats or the rangers) so they can fight against their attackers (like poachers or the potential terrorists).
Given these similarities, game theory would be a reasonable framework to handle these challenges.
TC: Can you briefly explain game theory?
FF: Yes! So we know that when we play rock paper scissors, we cannot always play “rock” because it can easily be exploited by the other players. We have to randomize it in some sense. Similarly, when we play poker games we cannot use the same action, and we have to randomize our strategy.
Game theory is basically asking how we can move beyond the games to model strategic behavior in real problems (like in policy economics) and how we can compute or prescribe the strategies that people, or agents would use in the end.
TC: What type of data do you need, for instance, in the poaching example? What type of data are you looking for from the rangers and how difficult is it to get that data?
FF: So for that case, the first thing we need to know is how the wildlife is distributed in the area. And then in addition to that, we want to know, the past patrolling information – where the rangers went for patrol in the past few years, and what they found along the way. Patrols will record where they found poaching signs, where they found other types of human activity signs, and those are all helpful for us to understand how the poachers behave.
TC: And do rangers typically record their patrols? Like, keep a paper record of where they went?
FF: Nowadays there are many conservation sites where the rangers have a handheld GPS with them as they walk and then they will ping down their location every few minutes, and also record what they find. So, we need that data. In addition, we need the geospatial features of the area – the elevation change, the land type, and the distance to the nearest village or distance to the nearest river or lake and so on.
TC: Does Google Earth have that kind of data to the resolution that you need?
FF: Google Earth can help in some sense, but it is not enough. We need detailed river information, and for some areas, the local government agencies have that record. And for the elevation, there is publicly available data. So that’s easy to get. And for land type, it’s hard for us to digitize it from Google Earth correctly. We need someone to tell us whether this is a forest and what type of forest it is.
Building upon this data, we can use machine learning techniques to predict where the poachers will place snares, but that’s not enough yet because the poachers also react or adapt to the patrolling strategies. If the patroller always goes to the locations the poachers used to go to, then the poachers will try some other locations. That’s where the game theory comes in. We need to consider how they impact each other’s actions and the figure out the optimal strategy in the end.
TC: That is so cool. What, in your opinion, makes a good problem to solve?
FF: Very good question. My opinion is that it is better if it is something that people are already doing. For example, the rangers have been patrolling the forests fighting against the poachers already. And the US Coast Guard has patrol boats trying to protect the ferries. What they’re not good at is optimizing the use of their resources. Optimization is at the core of artificial intelligence.
Once we figure out that it’s an optimization problem, it’s possible for us to figure out the best model for the problem and then try to find the corresponding solutions to it.
TC: What do you think children and parents can do to try to find a good problem that they can solve using technology?
FF: For children, I think it’s probably easier for them to find something where the model is quite clear, and they are familiar with it, meaning they don’t need to spend a lot of time figuring out the optimization goal. For the problems where they know what the optimization goal is, it’s easier for them to do it.
TC: Going deeper into the process of problem solving, what have been some really big difficulties you’ve encountered in your research and how did you overcome them?
FF: It’s quite challenging to find a balance between the theoretical contributions and the application. For example, in the wildlife protection project, we started with a simplistic theoretical model and then we derived some algorithms and showed some simulation results. But as we talked to the domain experts further, we figured out that the theoretical model does not have much use in practice because the model assumed you could go from area A to area B, but really the rangers are walking in hilly terrain. They have to climb over the mountains, walk in the river.
Then our collaborators invited us to go to the Malaysian jungle to do a trial patrol, and we spent eight hours walking through the jungle. From this we realized the theoretical models we developed in the lab didn’t make enough sense, so we needed to figure out a better way to model the real-world scenario. We created algorithms that can handle the actual terrain information and generate patrol routes that are compatible with the terrain. Moving from theory to practice and then back to theory is something that is really interesting and challenging.
TC: What inspires you to keep working and solving hard problems?
FF: I think first of all, curiosity. There’s always so many problems and you want to see how you can solve them. Here’s my feeling: there are a lot of societal challenges right now that can potentially be tackled by AI matters, ranging from poverty, food security to disadvantaged populations like homeless youths.
TC: How do you think AI can help poverty and homeless youth?
FF: One example could be in some developing countries the farmers want to sell their crops and they do not even have access to a huge highly-automated trading platform, and they go to their local markets and try to sell their products and sometimes the price is very low because the whole village is selling the same thing. In the last few years researchers from Canada (led by Prof. Kevin Leyton-Brown) helped develop a text-based trading platform, so you can text to this center saying I want to sell which product and at what price and then the center automatically tries to make matches between buyers and sellers so that definitely helps.
For helping homeless youth, researchers in the University of Southern California (led by Prof. Milind Tambe) design algorithms to improve the effectiveness of social intervention, by, for example, choosing people to receive access to homeless shelters so as to raise awareness about HIV among homeless youth.
My point is that there are many problems and there are many ways that AI can potentially be used to solve them, but we need more human resources like researchers and students and people in general to work on these problems who are trained to be capable of working on these problems.
TC: How do you think children can learn more about AI?
FF: I think it’s really important to make full use of existing resources, like YouTube channels and the Coursera type of online course platforms. And I think there are a lot of things that can be done, and we just need to be more open about all these possibilities.
So, for example if we have more YouTube videos introducing the basics aimed at teaching children, that would be super helpful because the impact can be greater with online courses or just online videos introducing the basics to the children than just inviting a few people to attend a series of college lectures.
TC: Right. And that’s exactly why we put our curriculum online for that reason. So last question – what new field of AI excites you the most?
FF: From the research side, I think I believe that the integration of machine learning and game theory is super important and promising. That’s why I’m working on it. And then from the application side I think there should be more work towards how we can use AI to solve societal challenges.
TC: Thank you so much.