5.10 Intellectual The Intellectual
Is the conflict between algorithms and social well-being natural? | Image source: pixabay.com
Introduction
Today, artificial intelligence algorithms have penetrated into all aspects of social life: shopping software relies on algorithms to judge consumers' consumption habits and push specific goods; ride-hailing software allocates driver capacity through algorithmic dispatch and pricing; and social software uses algorithms to analyze keywords, push advertisements and content. In this process, the algorithm that generates revenue for commercial enterprises as much as possible brings convenience to social life, but it is also criticized for privacy, discrimination and many other issues.
But is the conflict between algorithms and social well-being natural?
Fang Fei, an assistant professor at Carnegie Mellon University (CMU), believes that algorithms can also help solve some social problems by finding the right research direction. Her job is to combine game theory and artificial intelligence with real-world problems to solve complex problems in reality, such as protecting ferries from terrorist attacks, preventing poachers from harming animals in protected areas, and distributing food that is about to expire to those who need it.
Fang Fei's work is solid and steady, and he has won a lot of praise in the field. In 2020, she was named to the IEEE 'Top Ten Potential People in AI'; in 2021 she received the IJCAI (International Joint Conference on Artificial Intelligence) Computer and Thought Award; and in February 2022, she received the Sloan Research Award, an award for outstanding young scholars early in her career.
Fei Fang now works at the Software Research Institute of CMU School of Computer Science. She believes that in some cases, companies can also get higher profits when maximizing the total social benefits through algorithms. Compared with other strong application areas of artificial intelligence, AI is still a "blue ocean" for good. She hopes that more researchers will join the field and contribute to the overall social welfare.
She said that the most important thing to enter this field is to have the eyes to find problems.
The following is the dialogue between "Intellectuals" and Fang Fei, with the text deleted.
Written by | Wang Yiwei
Editor-in-charge | Chen Xiaoxue
● ● ●
Fei Fang, Assistant Professor at Carnegie Mellon University (CMU), and Sloan Research Award Winner when artificial intelligence is combined with game theory
Intellectuals: Game theory was proposed by von Neumann and Oscar Morgan Stern in 1944, when and how did game theory research combined with artificial intelligence develop?
Fang Fei: In the beginning, game theory was mainly in the field of economics (development), and Nobel laureates in economics often studied game theory. Due to the development of computers, everyone is very interested in what kind of problems in game theory are computable, or can effectively calculate the results, they will study different forms of incentives and different equilibriums, to see what kind of equilibrium concepts are effective algorithms in what kind of game, and what kind of problems are NP hard - more difficult to solve. With predictions in place, you start dealing with computational challenges.
The progress related to the security game came in 2006, when Tuomas Sandholm and Vincent Conitzer presented a paper at the Economics and Computation conference [1], saying that in Steinberg equilibrium, in a relatively simple problem, it is possible to solve the algorithm of polynomial time. But this problem cannot be solved when it extends to the payment function of a player in a game or when the payoff in the game has multiple possibilities and unknown types.
After that, my mentor, Milind Tambe, was the first to apply game theory to security. He studies whether a stackelberg game (a completely informative dynamic game in which there are two sides) can be applied to a specific problem, and each side chooses its own strategy according to the possible strategies of the other.
For example, we want to protect the airport in Los Angeles. In the issue of protecting airports, there are protectors and attackers. The rangers of the airport, the defensive side, patrol every day. Then as an attacker, you may spend a long time to observe what kind of rules the patrols have, whether they patrol Terminal One every Monday, patrol Terminal Two every Tuesday, and so on, and then find the relative weaknesses in the patrol plan according to these observed laws to avoid being caught by patrol personnel. They will have all kinds of attacks, for example, someone will bring pistols, rifles and other weapons into the airport.
On the one hand, patrol officers will set up checkpoints at multiple intersections entering the airport, and on the other hand, they will patrol the various terminals with police dogs, but because the number of checkpoints and the number of police dogs are not enough, they need to decide what kind of places to go and what to do every day.
Because of this asymmetry, they modeled the problem as a Steinberg game: the patrol side is the leader, first choose a patrol strategy; and the attacker side is the follower, observe the leader's strategy, and then respond, choose a better attack method to respond to the existing patrol plan.
On this basis, they developed a series of models and algorithms to study this field, and slowly became the first wave of security game projects.
After I joined Tambe's research group, I began to apply more machine learning methods. We began to study moving targets, such as whether ferries could be modeled with similar models, and whether we could find fast algorithms to solve the best way to patrol.
U.S. coastal defense patrols various crossings. In New York, part of the patrol's mission is to protect the ferry from Manhattan Island to Staten Island. In this problem, the ferry is the moving target, and the patrol is also moving in a small boat, and what we want to do is to use game theory modeling and build an algorithm to find the best route choice.
The original model was mainly a Steinberg game between the two sides, and later there were also many studies. In some of the animal conservation work of the first two years, people in the reserve told us that sometimes residents would tell them where poachers had gone or where they were going. At that time, the existing models could not take such factors into account, so we built a new model to consider the problem of more than two gamers.
Intellectual: It sounds very practical in this field. So what is the scientific goal of combining game theory and AI?
Fang Fei: In my opinion, its scientific goal is to find more efficient and faster algorithms to solve more complex or practical games. Game theory is the theoretical framework, and AI or deep learning is the tool to solve mathematical problems within this framework. In the whole field of computational game theory, what everyone is doing is hoping for better and faster algorithms that can solve more complex games.
Many times, a combination of computation and game theory fails to find a fast solution.
One case is that this game is too complicated to solve with mathematical programming, because if you use data programming to solve, it takes millions of electricity, millions of restrictions, if you really calculate on the machine, just to solve, the memory of this machine will explode. In this case, we can find a better solution to this big problem through deep learning, or a strategy that is closer to equilibrium.
In the second case, sometimes human behavior needs to be considered in the game problem. The most traditional game theory assumes that all people are rational people, but later it was proposed that not all people are completely rational, and it is necessary to understand the patterns of human behavior. Machine learning can help us model people's behavior through human past behavior data, and then go down to find solutions to other game parties in the game.
The third case is inversive game theory. The general game theory problem is to say, I tell you that the game is played by these few people, in different cases, what will be the return of each person, and then ask you what the equilibrium strategy should be for such a game problem. Reverse game theory is the opposite, I can observe what kind of action everyone takes in the game, and ask if I can find out what everyone's return function is.
In such a problem, there is a lot of observation data to describe each person's behavior and the actions he takes, and we have to reverse the benefit function from the action, and machine learning can also help us.
The three I mentioned are my own, using machine learning as a tool to solve problems in game theory. There are other ways to combine game theory and deep learning.
In turn, game theory can also be used to solve many problems in machine learning. One example is carnegie Mellon University recruiting students, and in the initial round of screening, everyone handed in the application materials and everyone came to see. In fact, in addition to people making judgments, the school has an algorithm to determine whether the student is worth considering, to prevent some people from being missed. These two lines are completely independent judgments, that is, as long as the man or machine thinks that the student should be worthy of consideration, then it will move on to the next round.
As you can imagine, in such a situation, if students know what kind of screening you are using, they may find a way to change some of their behavior or scores, and adjust their behavior according to the algorithm. In such cases, we can actually use game theory to analyze (how to solve this problem).
Intellectuals: What types of social problems can be solved by the combination of game theory and machine learning? How do you usually find these problems and pain points in real society?
Fang Fei: We helped the United States coastal defense to design patrol routes, and animal protection organizations to design rangers' patrol routes. The former is a safety issue and the latter is related to the sustainability of the environment. There are also mobility-related issues, such as transportation, which we are also studying.
Part of it is a coincidence of various factors, and part of it is actually because of some of the things we have done before, extending to similar or related problems.
For example, when I was in My First Year, I did the work of protecting ferries. After we gave the speech and sent the paper, we went to talk to different people, and someone mentioned that your question sounds very similar to the animal protection ranger patrol problem, and have you considered applying your algorithm to that problem? This gives us the motivation to talk to the relevant experts and understand the actual problems.
After I came to CMU, we also did a new series of work, which was food relief. I didn't know there was such a nonprofit in Pittsburgh at first, but I once met the CEO of this organization in our school conference room and exchanged business cards. I went to check what their organization was doing and found it particularly interesting. Later, I volunteered with them myself, and I went to talk to them, and told them that we did a lot of AI-related work, hoping to see what our technology can be applied to this problem and help your platform develop better.
Since the beginning of this discussion, there has been more and more real work slowly, and we have now sent three or four papers in this direction, and [algorithms] have been used in their systems, which is also very happy.
Food relief is divided into two parts, the first part is to match the donor and the receiver, that is, the food donor and the food recipient; the second part is to match the volunteer to deliver the food. For example, I was volunteering, a café in CMU had leftover cakes, they wanted to donate to the nearby single mother aid station, just on the mobile phone APP said that we need volunteers to send food from CMU's café to the aid station, I was very happy to see it, just upstairs, I quickly took the order and sent the food over.
Our main research is how to match the platform and volunteers, or how to help the platform better find volunteers who can deliver food, and how to let volunteers have a better experience in the process. In this case, our goal is not to send too many notifications, because if we send a notice to all the volunteers every time we come to the list (it will be very disturbing). We want to keep out too many notifications, but we can increase the acceptance rate and shorten the time.
After the new recommendation mechanism was launched, the overall order acceptance rate increased, and the average order receiving time decreased. This problem is actually a bit similar to the taxi software, and there is also a dispatch process, but what is very different is that this is a non-profit platform for pure volunteers, and the pricing strategy for the taxi software is completely unusable.
Find the real problem
Intellectuals: What do you find most difficult in finding a solution?
Fang Fei: It's looking for a problem. Find problems that can really generate value using AI.
Like the problem of food relief, after I started volunteering, I felt that there were too many points in the whole system that could be improved. I wrote a long email to their CEO, saying that in my experience, one, two, three, four, and so on can be improved, and then our AI can do something in each area. My suggestion is that your guidance information to volunteers is too vague, and you need to provide clearer guidance through some intelligent ways, such as who you should go to to connect with whom, where to meet others, etc., and you can let a volunteer send multiple orders.
Then I went to their office and talked to them, and they said that the questions you asked were interesting, but not what we really cared about. For their organizations, what needs to be seen is the bigger aspect, and their real pain point is how to get more volunteers to participate, so that food is not wasted, and it can be delivered to the people who should be delivered in a timely manner.
So to keep talking to them, my students also went to their internship for two days to help them do the dispatch work and experience what the dispatch process looks like from an internal point of view. In the continuous discussion, we finally slowly found a problem that is very concerned about them, and for us, artificial intelligence can really play a role. In my opinion, this is probably the most time-consuming step.
Another thing that I find particularly challenging is how to get them to actually use this thing that you make. Maybe he gave you the data, and then you made this thing that looked quite interesting and published the paper, but he felt, I don't have the ability to really implement your algorithm in the system, we are too busy, there are other things to do, this thing may not be their highest priority. If it's a purely business problem, this tool can increase your profits by 10 points, and they may be very active in doing it, but non-profit organizations are not like that, they have other considerations. How to persuade them to do practical tests and applications is not a very difficult point in our opinion.
Our experience is that first of all, we must fully respect their wishes and actively communicate with them; second, we must minimize the costs they need to pay, including time costs, labor costs, and possible monetary costs. When we are doing the actual testing, it is best to do all the work that can be done, they may only need to spend a few hours to have a meeting with us, discuss the plan, and then approve that the test before we go online is passed, and we can go online. That way they'd be more willing to do such a test.
Intellectuals: Your projects seem to be more public welfare now, taking into account the nature of the research itself, or do you think it is a work that maximizes efficiency?
Fang Fei: We are not rejecting commercial projects, but I personally hope to help solve some social problems. I joke that we are doing senior volunteers to help some government organizations or industrial non-profit organizations better provide higher quality services to society.
These issues are very important to society, and they may not necessarily produce direct business benefits, but these are problems that affect a lot of people, but not many people do. It is also understandable that now that artificial intelligence is so hot, it is conceivable that you can make a lot of money if you go to the industry, so when you have this ability, many people may choose to make money. Maybe a lot of people are interested in it, but not so many people are actually invested in it. I also hope that we can do some work ourselves on the one hand, and on the other hand, we can get more people to participate in this kind of work.
I teach AI for social good in school, and I hope to train more students to contact such problems and solve such problems.
Intellectual: As you just said, finding problems is a difficult point. How do you see the future development of this field, and can it be applied to more fields and more problems in the future?
Fang Fei: I think there are a lot of questions worth doing. I've been co-chair of the AI for Social Impact Special Track of AAAI for the past three years, and we've received a lot of papers every year, and I can see a lot of people interested in this and are constantly working on it.
For example, the 17 sustainable development goals of the United Nations, our current work may only involve four of them, in fact, there are many other goals, and there may be many other goals in the middle, which may also involve many issues worth doing.
Total social return vs corporate profit
Intellectuals: Many companies use AI algorithms to improve efficiency, but they have also been criticized by some. In takeaway companies, riders are declining because of the rules of these algorithms, or the strategy of the company itself. What observations do you have on relevant AI applications, and what can be done in terms of corporate responsibility or the use of AI algorithms to improve?
Fang Fei: Companies must pursue profits, which is definitely their main goal. In addition to profit, game theory or mechanism design often talks about a social welfare, social welfare, and the sum of the benefits of all.
In a takeaway platform or taxi platform, the income of the platform plus the rider or driver, as well as the people who wait for meals and the people who want to take a taxi, the sum of the income of all people is a objective function that needs to be paid attention to. When we were doing research on the pricing strategies of ride-hailing platforms such as Uber, we set the goal at maximizing social benefits. How is social gain calculated? For example, for the passenger, he wants to go to this place, he is willing to spend 100 yuan to go, and then he pays 50 yuan, then his income is 100-50.
It's hard for us to tell commercial companies that you don't have to focus on profits, it's very difficult to focus on the total social gains. But what we see is that in what cases people will pay more attention to the goal in addition to profit.
One is regulation, such as insurance companies, may be at the beginning, when there is not enough supervision, the pricing strategy of the insurance company will be very biased, when some things are exposed after the regulatory intervention, the supervision may require your insurance pricing strategy to be more fair. In such cases, then, these companies will consider the total social benefits in addition to profits, such as fairness.
The other is that, in some cases, we can go and prove to these companies that, in your case, maximizing total social revenue is actually very small a difference from maximizing total profit. If you maximize the total social benefits, you can also get a relatively large benefit, which may not necessarily be maximized, but it is also relatively close, but it can have a greater improvement in the overall social welfare. When we studied the taxi platform, we did an analysis, and under certain assumptions, we can also get relatively high profits when maximizing the total social returns.
How can data bias be eliminated?
Intellectuals: In AI research, data is the foundation, but there are often cases of inaccurate or biased data, will your research encounter similar problems? How to solve these problems?
Fang Fei: Yes. In the case of animal protection, the data we collect has a lot of problems. For example, the existing data is collected by rangers who have worked very hard to patrol and collect. But they do not cover all the areas, maybe some areas go more, other areas go less, which leads to more areas to go, there may be more data, and how high the incidence of poaching in this place is more accurate, can have a relatively accurate estimate; but for those places that go less, even if they say I went, I did not find a hunting set, does not mean that there has never been poaching in that place, it may be just because they have gone less, When I went there, I didn't find a hunting set.
And if you look at the overall amount of data, there are definitely fewer points where they have found hunting sets in all the places they have visited, or there are more times when they walk on the road and find nothing. This also illustrates the imbalance in the data and the challenge we need to deal with when designing machine learning algorithms to learn the behavioral models of poachers.
We tried a variety of methods, like the original version of the algorithm, we divided the entire protected area into multiple plots, if the amount of data in this plot is relatively large, you can use a slightly more complex machine learning algorithm; if the amount of data in this plot is not enough, we will combine all the data of those plots that do not have enough data, and then find a machine learning algorithm that does not have so high data requirements, such as decision trees and other methods to predict. That was the idea at the beginning.
Later, we made a lot of other attempts, such as in the Huangnihe Conservation Area (located in Jilin Province), we sent a questionnaire to the local rangers, asking them whether the overall poaching risk in each area was high or low in the multiple areas of the entire protected area we divided, and according to their answers, we went to sample some additional data points and put them into our data set, which on the one hand helped us add more data, and on the other hand, we could correct the bias of these rangers themselves.
Has artificial intelligence entered a dead end?
Intellectuals: There's been a recent argument that AI has reached a dead end and is only suitable for dealing with questions that have low risk and perfect answers. What do you think of this statement?
Fang Fei: On the one hand, we have to admit that artificial intelligence is still not so advanced, it is really not so powerful. But I think there's still a lot that can be done.
One is the high-risk problem. Relatively speaking, in high-risk problems, artificial intelligence is to assist human decision-making, not to replace human decision-making, we know that decision-making is high-risk decision-making, what we do is to hope to provide decision-makers with more information and more options.
There are a lot of people doing interpretable AI now, in part to solve the problem of applying AI in high-risk scenarios.
One of the directions is that in the end I don't want a neural network, I need a decision tree, a rule-based classifier that can be drawn and directly understood by people. Maybe it doesn't perform as well as deep learning, but it can still perform better. In the process of training decision trees, I may still need deep learning, but the final presentation is a decision tree, so that people can at least understand.
There are other routes, such as Whether I predict or make decisions or use deep learning models, but I explain to people in natural language which characteristics play a key role in making such predictions and decisions, so that it finally makes such predictions or decisions.
There are various other directions. We are also doing some of this ourselves, hoping to open some of the black boxes in artificial intelligence, so that the relevant decision-makers can really understand what AI is doing, at least to check and verify whether these things found by artificial intelligence are useful or not, and then decide whether to use it or not.
Of course, on the other hand, I think that the AI for social good has not yet reached the bottleneck period, and there are still many problems worth doing, so I hope that more people can pay attention to this aspect of research and be more willing to do some research in this area.
When you say that many problems have been solved, it may be problems such as image classification and medical image recognition. Many people have been doing these problems for a long time, and have reached a very good but difficult place to go further. But I think AI for social good this piece is still the state of the blue ocean, because relative to image processing speech processing or natural language processing, there is no standard data set or a specific problem, we must continue to develop better and better, more and more new algorithms, to understand different problems, what is in this problem is what AI can help solve, what kind of method is the most suitable to solve, and finally how to do systematic testing, promote landing. In addition to algorithm design, there is a lot of other work to be done.
Intellectuals: Is there any direction you want to do in the next 2 or 5 years, or what work you want to do?
Fang Fei: The areas I hope to dig into now are some issues related to animal protection, food rescue and transportation.
One ongoing is our ongoing work on animal conservation in collaboration with the World Wide Fund for Nature (WWF). We want to help them automatically collect news reports and government reports, find articles related to nature reserves and animal protection, and then organize those articles into a visual form that can be directly viewed and analyzed, which can save them time. They are now doing these things by human beings.
This work has been field-tested and is already being used in WWF's internal systems, and we are constantly gathering feedback and improving our work. This problem does not actually involve the problem of the game, but involves a lot of natural language processing.
In the food rescue area, we still hope to be able to put the existing algorithm on the ground, and then explore new problems.
We also have some work in progress related to cybersecurity, and there is clearly a game in this issue. We have done some game theory models and algorithms before, and now we still want to derive better, closer to the actual model, better algorithm. We also want what we do to be closer to the stage of being actually used.
Intellectual: I have a perhaps unrealistic idea, thinking of the recent epidemic in China, everyone has a lot of hesitation in getting vaccinated. Is it possible for your work to solve the problem of encouraging everyone to get vaccinated?
Fang Fei: If there is already some reward mechanism, we may be able to analyze this mechanism, maybe we can reward it more finely, so that the final effect is better, which is a possible direction.
Advice for learners
Intellectuals: What advice would you give students who want to learn the direction of combining game theory and deep learning?
Fang Fei: One is the foundation or the foundation. At the undergraduate level, if you have access to courses related to game theory, as well as deep learning, multi-agent related courses, it is still recommended to study, especially if there are some course projects, or projects that you are more interested in, sometimes the process of doing projects can help you better understand what you have learned. This project doesn't even necessarily mean that you're going to end up with a paper, it's probably about doing some exploratory work that interests you.
Intellectual: Finally, I would like to ask, how did you feel about winning the Sloan Award?
Fang Fei: I really didn't expect it, because I applied once before and didn't get it. I applied again this year, but I didn't actually have high expectations, and I felt very honored to get it later. Because it is not only for the computer field, it is for scholars in many fields, so it seems to get more attention.
In addition to my own award, which surprised me, another thing I think is quite surprising, is that there are a lot of Chinese or Chinese in the list of winners this year, and there are many girls, which may not have been seen before, indicating that we (Chinese) are doing a good job in North America.
Resources:
[1] V. Conitzer, T. Sandholm, Computing the Optimal Strategy to Commit to, EC’06, June 11–15, 2006, Ann Arbor, Michigan, USAhttps://users.cs.duke.edu/~conitzer/commitEC06.pdf
Plate editor| Ginger Duck