Shanghai Fortune University Professor He Simai: Application of AI in Kidney Transplantation

Release date: 2017-08-01

On the afternoon of July 20th, the "AI Masters Roundtable Series Conference" jointly sponsored by Shanji Technology and Titanium Media was held at Howard Johnson Plaza Hotel Shanghai. During the meeting, Dr. Tang Pingzhong, an assistant professor at Tsinghua University, first introduced the combination of artificial intelligence (AI) and economics.

Later, Prof. He Simai, a professor at Shanghai University of Finance and Economics and a recipient of the Young Technology Award of the China Operations Research Institute, introduced how AI can be combined with operations management and business practices.

Professor He used kidney transplantation as an example to explain how AI can help kidney transplantation improve matching efficiency. With the help of current AI system, the efficiency of organ exchange transplantation matching system has been improving, and the best system efficiency has reached 90%. It is to ensure that the theoretical optimal solution is at least 90% matching rate. In addition, the number of kidney transplants has increased from 28,000 to 33,600 per year in various non-profit kidney transplant institutions.

Later, Professor He also introduced how AI can help merchants to price. Pricing is a very dynamic and fast-changing process. Customers, friends, and the nature of the goods will cause significant changes in pricing, and AI will be required to help the business make important decisions.

Dr. Hesmai gave a speech

The following is the full speech

Hello everyone, just now, Tang introduced how game theory is combined with AI. Next, I mainly introduce how AI combines with operational management and business practice.

The speech is mainly about business intelligence, which is artificial intelligence in operations management. The main content is divided into three parts, the first part is the application of AI in organ transplantation, and also pays tribute to Professor Sandholm (he is an expert in kidney transplantation). This series of work is of great significance to the public interest because of the extreme needs of society.

Another guest on the scene, Professor Tuomas Sandholm, the developer of "Cold Master"

The second part is about revenue management, which focuses on how to combine AI with a specific business environment when it comes to specific decision-making issues such as pricing and inventory.

In the third part, when it comes to revenue management, including inventory management, it also needs to consider some strategic customer behaviors, which involves a strong game background. When customers observe or guess the decision-making rules of the merchants, they will make some strategic purchase behaviors, which will have a great impact on the merchant's strategic effects and optimal decision-making.

Artificial intelligence applications in the field of operations management are mainly divided into three major blocks. Part of it is very close to the traditional AI, that is, how to analyze the data after having data, and how to rely on the data for prediction. Part of it is a model that needs to be tailored to the specific business environment and needs. Part of it is algorithms, because business scenarios often require complex decision-making problems within limited time and computing resources.

Traditional AI now has some great applications, such as the famous Alpha Go. In the fields of image recognition, speech recognition, and text translation, AI has done a good job. These areas have a common common feature and the goal is very clear. For example, image recognition, face recognition, judging criteria is how high the recognition accuracy? Generally speaking, both the problem and the target definition are relatively clear, and the external environment is relatively stable.

But in business intelligence, we face completely different problems and environments. Today's business environment may not be the same next year, the economic environment may change, and competitors are changing. Especially in the field of application of AI, many applications are currently related to e-commerce. But for mainstream e-commerce, the current annual growth rate is extremely high. Perhaps every year, its product structure and customer structure are changing.

The second big problem is that there are many goals in the business environment, but there is no single clear goal. Maybe the company is inside, and some departments want to pursue the optimization of inventory management. Other departments want to ensure the order satisfaction rate, the higher the better. There is often a balance between sales volume and profit in revenue management. How should we make decisions when the goals pursued between different departments are not consistent? In the specific actual scenario, if the difference in demand between different departments can be fully explored, the complex relationship between different constraints can be fully combed and considered, and the improvement of income is very significant. Therefore, it is necessary to combine the problems of the AI ​​technology and problems through the model itself.

AI and organ transplantation

Here are some specific examples. The first is organ transplantation. Traditional organ transplantation is mainly two kinds, one is the donation by relatives. Another traditional form of donation is the donation of the deceased, and the deceased signed an organ donation agreement before he died. However, basically any kind of organ, the amount of donation from the deceased is currently far from enough. More serious kidney transplants, only 10%-20%, this number is very very low, then what should I do?

In medical practice, there are often relatives who are willing to donate but cannot donate, often because of blood type and various antibodies. Therefore, there may be such two families, A's family can not donate to their relatives A, but is suitable for another patient B. The organs of B's ​​relatives are suitable for patient A. When this happens, both parties can solve the problem by exchanging donations. This application is becoming more and more popular, because it can save many lives. Later, everyone found out, can you do more? For example, I can't match two people, but the three families match in a circular way, and four families and five families can.

The most historical record is that at the same time more than a dozen families constitute a large cycle of surgery, but this transplant program has a serious problem, that is, multiple operations must be done at the same time. Because kidney transplants may cause donors to suddenly repent, this is a serious problem, especially when the relatives of the donors of the counter-official party have already donated. In order to avoid these problems, the current law is strictly regulated and must be exchanged at the same time. But even then, there are still some problems.

Multi-person exchange kidney transplant

Another method is chain matching. To put it simply, the organ donated by the deceased is transplanted to a patient, but only if the relative of the patient promises to donate the kidney to the next patient. The relative of the patient promises to donate the kidney to the next one to form a chain reaction. The biggest advantage of chain transplantation is that it does not need to be performed strictly at the same time. In every family, the patient first receives the kidney and then the donor donates the organ. If someone repents, although there will be some negative effects, it is relatively small. Another benefit of chained migration is that it is highly selective and, relatively speaking, greatly increases the efficiency of system migration. The disadvantage is that it will consume a portion of the kidney source donated by the deceased.

Chain transplantation of the kidney source of the deceased

At present, the United States has officially enacted legislation in 2007 to allow exchange of transplants. Some institutions and platforms have been established as kidney transplant exchanges, such as UNOS. Primarily non-profit, the goal is to collect data from patients and data on relatives who are willing to donate organs to better match these relatives. After 4-5 years of establishment, the number of kidney transplants has increased from 28,000 to 33,600.

However, chain transplantation will face an ethical problem, and the kidneys donated by the deceased are far from enough to meet the needs of patients who have not been donated by relatives. A kidney can save a lot of patients by chain transplantation on the one hand, but at the same time there must be patients who cannot get a proper organ transplant. This is a bit like the famous tram puzzle, where a group of children are playing. When you see a train coming over, you can choose to take the train to a ramp to save the children, but it will hurt another person working on the ramp.

Famous tram problem

This is a universal moral problem. Therefore, when kidney transplantation, the decision is not simple, the more people, the better. Because some types of patients, because antibodies and blood types are particularly difficult to match, this special patient will be backlogged when only the short-term transplantation efficiency is the highest. At this time, it is not a simple prediction. I have to predict how long each kidney can live after being donated to a patient, and how high the survival rate between them is. At this time, we use machine learning to make artificial intelligence decisions, and the efficiency will be much higher.

This kind of decision is a typical multi-stage dynamic decision, involving some conflicts. How do you solve this situation? At present, the advantage of using the integer programming algorithm in UNOS is that the porting rate is relatively high, and the number of people in each stage is relatively large, but there are also some problems. We recently had a paper that proposed a relatively simple random approximation algorithm, which can be used to achieve high efficiency even without relying on a computer. And difficult patients will not have a backlog. At present, the efficiency of this organ exchange graft matching system has been increasing, and the best system efficiency has reached 90%, that is, the matching rate that can achieve at least 90% of the theoretical optimal solution.

It is worth mentioning that the study of these decisions, theoretical in-depth discussion, is very difficult. Professor Roth, the Nobel Prize winner at Stanford University, has done an outstanding job. Professor Sandholm, who is present today, is the actual algorithm manager of UNOS, the largest source of kidney reservoirs in the United States. In the past few years, we have been trying to answer this question, especially the estimation of effective algorithms in the actual network, and have made some theoretically unexpected breakthroughs, which have received the attention and praise of Professor Sandholm and others. It is also the main reason why we first met him and were able to invite him to lecture in Shangcai.

Application of AI in revenue management: merchant pricing problem

Next, let's introduce the application of AI in revenue management. This is based on the actual problems we encountered when working with many home appliance companies. In the middle of revenue management, one of the key issues is how to price. This problem is very complicated.

First, the pricing logic for different products is different. Many merchants simply predict how many pieces a price can sell and then choose a price that looks more appropriate. But this is not enough. Some products use the simplest cost pricing in the market, which is the cost, plus a weight, to generate the price. In the e-commerce environment, the pricing logic adopted by most key products is that the price is alive and the price is guaranteed to be lower than the price of the competitor. Everyone can now see that many e-commerce platform products will guarantee lower prices than competitors' products. If they are high, they will compensate for the losses. There are also some products, the competitive environment it faces is not very intense, then its pricing logic should consider the relationship between supply and demand.

This time is often a relatively large profit margin. There are other pricing methods, such as price discrimination. Different from general pricing, it is a uniform price for all customers. Price discrimination needs to give different prices to different customers, thus maximizing profits. We also propose a pricing model called robust pricing. Essentially, even if the prediction accuracy is not satisfactory, it is still possible to guarantee the pricing strategy given by the model and theory, and the risk is controllable. In summary, if there is no good understanding of the product, then it is impossible to achieve a proper pricing. For different products, we must first understand the different product positioning and determine the basic pricing logic.

Secondly, pricing needs to consider very complicated factors. When simply analyzing data, it is easy to miss some key factors. For example, you need to consider the price of your friends and the psychological price of your customers. Today I have been pricing 20, but the friend may be adjusted from 20 to 25. If the data does not include the price factor of the friend, it is impossible to make a high-precision forecast.

In addition, there is the sensitivity of the customer price. Some customers don't care about the price. When adjusting a small amount, he cares about your service. But some other customers will be more sensitive. This will create an interesting question. Is the business doing periodic promotions, or is it always keeping the price steady and better? There are also seasonal products. Or alternative products, such as mobile phones, Samsung, Apple, Huawei, etc., have serious substitution effects between them. Finally, there are complementary products. When customers buy computers, they will buy keyboards and mouse accessories.

In 2015, "Double 12", a shopping mall in Zhengzhou also played a promotional sign @视觉中国

The biggest challenge when pricing is promotion. Not only do you make promotions yourself, but competitors also make promotions for a long time. Many products, you will find that 90% of the time, you or your competitors are doing promotions. How to price at this time, the promotion will have a variety of effects.

For example, what is the use of objects from the price and price comparison? Products with high attention and high competition have a serious problem at the time of price comparison. To calculate the customer-to-price relationship between price and price, traditional statistical or economic models often use parametric models, and many people use linear models. However, at the time of price comparison, the volume-price relationship is generally extremely disconnected from the linear relationship. Customers are sensitive near the psychological price segment, and are less sensitive when they are far from the psychological price. There are many problems involved here, such as how to capture the psychological price, which is often related to the price of friends and the price of similar products.

Another one is that I adjusted the price, and the friend is likely to follow the tune. This situation is dynamic. I can choose to adjust the price with my friends. I can also avoid the price adjustment and do the promotion, and get it when he does not make promotions. The tools to be used at this time include web crawlers and machine learning to predict and capture each other's behavior. Because of the game involved, we need to know in the market, we are a price leader, or a price follower, what is the relationship with the competitor? Specific decisions and market positioning are closely related.

In addition, what is the balance between my current interests and long-term interests? For example, Amazon's strategy is to maintain absolute low prices among competitors. This strategy has been a great success in the United States, making Amazon a monopoly. But the same strategy in China is very poor, and Amazon is basically squeezed out of the market.

In the US, Amazon's online sales accounted for 14.3% of total sales (including offline sales), far exceeding the second Wal-Mart online store (Photo: statista)

With regard to the relationship between supply and demand, we can see that with the supply and demand relationship to make pricing, through an overall adjustment, sales, revenue and profits can be greatly increased, and the range is quite high, which is the effect of robust pricing. The figure shows the effect of the ordering strategy of a flash purchase e-commerce platform. We adopt two stages, the first stage adopts robust ordering, and the second stage combines artificial intelligence learning means to apply new data, which can generate great performance. Increase in range.

Source: Observer Network

Recessed LED Panel With Back Light

Recessed LED Panel With Backlight,Backlit Recessed Round LED Panel Light,LED Square Panel light

Ningbo Wellway Optoelectronic Technology Co.,Ltd , https://www.wellwayopto.com