Introduction to the series
As a shared travel company, Didi uses information technology to build a real-time and intelligent online trading market . In this huge market, Didi adheres to the principle of putting user value first, and constantly improves its technology to achieve more efficient transactions. Operational efficiency and a more intimate user experience.
In order to enable everyone to understand the online trading market and its technical challenges, expand technical horizons, and enhance technical exchanges, this series of articles is shared to fully introduce the main areas of trading market strategies and share existing exploration experience for readers.
Although this series of articles involves more professional technology, the writing strives to get an introduction, and is aimed at friends with various professional backgrounds, especially friends with knowledge backgrounds in the fields of computer, logistics, and transportation.
Through this series of articles, you can get:
1. Understand the core elements and core issues of the trading market, and outline a panoramic picture of the trading market
2. Understand the commonality and characteristics of Didi as a travel transaction market, focusing on the main areas involved in online car-hailing transactions
3. Understand the major areas and technological developments in trading strategies, including:
a. Driver matching
b. Supply and demand adjustment
c. Behavior recommendations
This article is divided into:
1. What is the trading market
1) The nature and network effect of the trading market
2) Improvement of social efficiency brought about by the trading market
2. What are the technical characteristics of Didi Trading Market?
1) Mechanism design
2) Decision intelligence
3) Operations Research
4) Reinforcement Learning
5) Causal inference
3. Summary
1. What is Marketplace
Trading was born along with human civilization. From the ancient bartering to the modern service trade, effective exchange behaviors have promoted the continuous advancement of human society. The trading market is an environment and rule system for buyers and sellers to conduct transactions.
The modern trading market not only provides the function of buying and selling goods or services, but also builds a stronger trust system and communication mechanism for buyers and sellers through platform capabilities. Through open, transparent and fair trading rules, it protects the vital rights and interests of buyers and sellers and improves the entire market. The operating efficiency of the environment can reduce the social loss caused by the misallocation of resources.
With the development of transportation and communication technology, the trading market is no longer limited to a specific geographical location, but can connect buyers and sellers around the world through the Internet. In particular, the popularization of computer technology, especially the application of a new generation of artificial intelligence technology since 2006, has also enabled the trading market to conduct transactions more quickly and accurately, achieving more efficient, more considerate and more reasonable market trading decisions, and further Improved transaction efficiency and experience.
The nature and network effects of trading markets
The essence of the trading platform is connection and matching. It hopes to connect the supply side and the demand side of products or services that are difficult to contact in order to achieve matching and matching, thereby improving the efficiency of the entire market operation.
It can be seen from the above figure that the more bilateral participants in the trading market, the more obvious the network effect will be, namely:
1. Let the suppliers of products or services face more potential demanders and select more desirable buyers from them;
2. Allow the demand side of products or services to face more suppliers and their differentiated products or services, and choose more desirable products or services from them;
This network effect provides more possibilities to improve the efficiency of the entire trading market. In this regard, academia, especially in the fields of economics and industrial engineering, has also done a lot of research.
The online two-sided market in the Internet age expands the supply side and demand side to communities, cities, countries and even the world, greatly improving the network effect (Network Effect), bringing greater market efficiency and experience for the whole society upgrade.
Network effects bring multiple benefits to the development of the entire trading market:
· As the number of participants increases, the network connection increases at a quadratic rate, and the scale expands rapidly
Larger two-sided market provides more and better choices for both parties
The more historical transaction behaviors, the platform can better predict future transaction behaviors and provide better services
The improvement of social efficiency brought about by the trading market
Due to the blessing of computer network technology, information transparency and real-time communication have made the efficiency of the entire trading market unprecedentedly improved, further providing lubrication and improvement for the operation of the entire society.
In 2017, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) developed a new set of shared car dispatching system, and conducted a simulation test, which proved the effectiveness of shared cars in treating urban traffic congestion and improving social efficiency. Excellence.
The figure below shows how to match and dispatch drivers (green line) to meet passenger demand (asterisk) in this method. Through simulation research, it is found that this method greatly improves the overall traffic operation efficiency.
图片来自：Alonso-Mora, J., Samaranayake, S., Wallar, A., Frazzoli, E., and Rus, D.: On-demand high-capacity ride-sharing via dynamic trip-vehicle assignment, Proceedings of the National Academy of Sciences, Vol. 114, No. 3, pp. 462–467 (2017)
2. What are the technical characteristics of Didi Trading Market?
The emergence and rapid development of the Internet has brought earth-shaking changes to the trading market. Nowadays, the trading market can be simply divided into two categories: the online bilateral trading market for goods and the online bilateral trading market for services.
The online bilateral transaction market for items is represented by Alibaba, JD.com, Amazon, eBay, etc. These trading markets connect buyers and sellers through the Internet platform to realize a series of processes such as product display, transaction, payment and delivery. Most of these platforms provide a wide range of product categories, from ordinary consumer goods to high-end luxury goods can be found on these platforms. Of course, these platforms also face responsibilities that need to be borne, such as trust issues in product quality, handling of transaction disputes, and so on.
The service online bilateral transaction market is represented by Didi, Uber, Airbnb, etc. These platforms provide online transactions of services rather than commodities, such as travel services, accommodation services, and catering services. These trading markets use the Internet platform to connect service providers and service demanders, providing users with a more convenient and flexible service experience. These trading markets also strive to ensure service quality and address service security issues.
As an online bilateral transaction market for travel services, Didi provides a transaction platform for travel services for the whole society. This platform brings together hundreds of millions of travel users and tens of millions of drivers, and builds a huge bilateral transaction market. Experience many excellent, diversified and reasonably priced travel products.
In the huge online service bilateral transaction market built by Didi Chuxing, there are the following characteristics in terms of technology:
1. Combination of online and offline : Passengers issue orders online, drivers receive orders online, and then complete travel services in the real world;
2. Fairness and security of platform rules : the platform provides a transparent and open transaction environment for all passengers and drivers, and makes continuous efforts to ensure the safety of both parties;
3. Data security and privacy protection : Didi needs to ensure the security and privacy of personal data of drivers and passengers, and adopts a series of technical management measures to ensure data security and privacy;
4. The experience of drivers and passengers : Passengers hope for faster pick-up service and cheaper prices, drivers hope to obtain stable and reasonable income, and the platform needs to work hard to ensure the experience of all parties;
5. Diversified service forms : Didi provides a variety of service forms including express train, preferential enjoyment, private car, luxury car, taxi, etc., to meet the different needs of different groups of people for travel services;
6. Large-scale global decision-making problems : When faced with such a large-scale transaction problem, a large number of traditional computer algorithms have failed, and it is necessary to actively introduce artificial intelligence and big data technology to pursue more intelligent intelligent decision-making capabilities;
7. Reasonable scheduling and route planning capabilities in the smart city system : through digital twin technology, simulate the real environment in the virtual environment, predict and dispatch travel resources in advance, plan driving routes, improve traffic utilization, and reduce congestion;
8. Autonomous driving technology : When facing the future autonomous driving world, actively explore autonomous driving technology to improve travel safety and service quality while reducing operating costs.
How does Didi's trading strategy work when faced with a huge and complex trading market?
Mechanism design
The 2007 Nobel Prize in Economics was awarded to three outstanding scholars, Leonid Hurwicz, Eric Maskin and Roger Myerson , in recognition of their foundations for mechanism design theory.
Mechanism design (also known as inverse game theory), as an important branch of economics and game theory, aims to solve problems in the market environment by designing economic mechanisms or incentives to achieve specific goals. Mechanism design theory has applications in a wide range of fields, from economics and politics such as market design, auction theory, and social choice theory, to technical fields such as Internet interdomain routing and search auctions.
The optimization of the Didi trading market strategy is a broad mechanism design, which aims to design an appropriate trading mechanism based on the characteristics of the market and the needs of participants, so as to achieve the goal of win-win for all parties and promote the benefits of both supply and demand.
Decision Intelligence
Decision-making intelligence in a broad sense is an engineering discipline that combines social science, decision-making theory, management science and other theories with AI and data science. It helps people use data and algorithms to make decisions in complex social systems, thereby improving life, work and life. the world around.
Discussing decision intelligence in the trading market can be narrowly understood as using operations research, machine learning, causal inference and data science to make better trading decisions. A few simple examples can be given:
1. Make suggestions for passengers. When passengers have individual needs, they can choose which models and prices to get the right car;
2. Make suggestions for the driver, suggest better decision-making behaviors for the driver, plan the optimal cruising route, and achieve a better driving experience in the process of dispatching the car throughout the day;
3. From the perspective of the platform, when the rainstorm is predicted, in order to more fully meet the passengers' sudden demand for taxis in the future, from the perspective of the current situation of the whole journey and future predictions, it is decided to dispatch more drivers from a certain direction;
"Decision intelligence" is evaluated by Gartner, a well-known technical consulting company, as one of the emerging and important technologies that are in the period of rapid technological growth.
In Didi’s trading market environment, in the face of complex decision-making issues, we usually need to make multi-faceted decision-making work, including but not limited to: real-time matching decision-making, supply-demand adjustment decision-making, passenger-driver recommendation scheme, dynamic decision-making, safety and judgment Responsibility decisions, growth decisions, and so on.
Operations Research
Operations research is an important branch of modern applied mathematics, and has a wide range of practical applications in production and management. Facing complex problems in real life, operations research uses statistics, mathematical models, algorithms, etc. to seek optimal or near-optimal solutions, so it is also called optimization theory.
Traditionally, the scope of operations research research includes: mathematical programming (linear programming, nonlinear programming, integer programming, stochastic programming, combined programming, etc.), graph theory, network flow, transportation problems, network planning, queuing theory, storage theory, Game theory, search theory, decision analysis and other fields. With the arrival of the Internet of Everything era and the improvement of computing power, operations research is showing a new charm.
For the complex scenes of Didi, almost all problems have realistic complexity, such as: optimal solution under a lot of constraints, driver queuing, resource allocation, etc., all of which will rely on formal mathematical modeling of operations research to help us solve problems better.
Of course, operations research is not a "silver bullet" (Silver Bullet). We also fully understand that operations research, as a very long-standing subject, has many practical difficulties in its implementation and development:
On the one hand, abstracting real-world problems into mathematical programming problems and quickly solving them through large-scale solvers is a powerful weapon for us to pursue the optimization of business effects, and it also exposes us to many unknown areas of exploration.
For example: after the user opens the relevant permissions, in order to avoid disturbing the user too many times, limit the total frequency of recommending preferential information to the user (constraint condition), limit the interval of push (constraint condition), and finally make the preferential information obtained by the user optimal ( optimization goal).
In particular, when there is no unique optimal point, the entire income surface will constitute a "Pareto front", and we can choose different points on the Pareto front to better meet the various needs of users (multiple Target optimization), and through the improvement of product power, the Pareto frontier can be promoted as a whole, and the effect can be improved.
Literature:Siurana, Maria & Fernandez de Cordoba Castella, Pedro & Montagud, Arnau & Reynoso-Meza, Gilberto. (2017). Modeling and multi-objective optimization for simulation of cyanobacterial metabolism.
On the other hand, machine learning—especially deep learning—brings new opportunities and challenges to operations research. In the development of traditional operations research, more emphasis is placed on mathematical derivation and heuristic methods, while the powerful learning and computing capabilities brought by deep learning are continuously injecting more energy into operations research. Most of the large-scale operational research problems we face are difficult to rely on traditional pure mathematical methods to solve, and need to be combined with deep learning methods to achieve fast approximate solutions or better representation of constraints.
Reinforcement Learning
Reinforcement learning is the third basic learning method alongside supervised learning and unsupervised learning in the field of machine learning. It does not require labeled learning data, but through continuous exploration of itself and the environment to find the maximum return. decision-making method.
Let's think about it, when you need to make action decisions continuously in the face of the environment, if you take the decision with the highest current benefit at each step, you can get the best of the entire process, you don't need to use reinforcement learning, you only need to learn single-step The best way will do.
The problem to be solved by reinforcement learning is that in this type of sequential decision-making problem, the decision-making at each step needs to "leave a way out" or "create more opportunities" for subsequent actions, so as to make the whole better.
Take the Didi scene as an example: a driver in the semi-assignment mode (allowing optional orders) may consider "take one order to the hot zone, there will be more orders there in the future", in this way of thinking , each order may tend to "create more possibilities for the future", making it easier to achieve higher income throughout the day.
The DiDi platform uses reinforcement learning to solve a wide range of problem scenarios. Among various sequential decision-making problems, the Environment, State, Action, and Reward of reinforcement learning can be constructed from different perspectives, and different algorithm results and strategies can be obtained, thereby helping Drivers and platforms have greater success in a range of sequential decision problems.
Causal Inference
“Correlation does not equal causation” — this is a point that Judea Pearl, 2011 Turing Award winner and AI pioneer, keeps calling attention to.
In the late 1980s, Pearl mathematized the causal model, which changed the understanding of statistical causality in social sciences. Together with Professor Michael I. Jordan from California, he combined Bayesian networks, Markov Networks, probability methods, etc. are introduced into artificial intelligence, and a series of probabilistic map machine learning models have sprouted.
Regarding "correlation does not equal causation", a simple example is: in recent years, the level of carbon dioxide in the atmosphere and the level of human obesity have continued to rise. This correlation will make people misunderstand that carbon dioxide causes obesity, but the fact is that economic growth Humans eat more food and get fat, and economic growth also produces more carbon dioxide.
Many times what we want to study is: when a certain behavior (usually called Treatment) is changed, what is the impact of a certain variable that we want to affect through causality (usually called Effect). Moreover, in many cases there is no way to do a completely random comparative test to get an accurate average impact.
This research has expanded to many fields: the 2021 Nobel Prize in Economics was awarded to David Card, Joshua Angrist, and Guido Imbens, in recognition of their establishment of a new framework for causality research that can be analyzed using natural experiments and has revolutionized society Empirical research in the field of science.
In Didi business, drivers and platforms will make different decision-making choices, resulting in different business results. If we want to make better decisions, we not only need to understand "correlation", but also "causality" , through causal inference modeling to achieve accurate prediction of the consequences of the decision-making choice, and then provide better choices to protect the rights and interests of drivers and passengers.
3. Summary
As a domestic travel platform, Didi has built a complete online trading market to provide high-quality trading services for both drivers and passengers around the clock. In the more complex trading market environment, the greater the challenge of market decision-making.
In the complex trading market, Didi continuously improves market efficiency and driver experience through technological progress, constantly overcomes various technical challenges, establishes fair platform rules, provides comprehensive security guarantees, and provides diversified market service forms. Optimize the efficiency of transaction decision-making and achieve a win-win situation for all parties.
The above content briefly summarizes the technical challenges faced by the Didi trading market. I hope the above content can help you understand the trading market. In the future, it will bring specific technical content about driver matching, adjustment scheduling, and behavior recommendation.
4. References
1.Evans, David S., et al. "Platform Economics: Essays on Multi-Sided Businesses." (2011).
2. Pratt, Lorien. Link: How Decision Intelligence Connects Data, Actions, and Outcomes for a Better World. Emerald Group Publishing, 2019.
3. Myerson, Roger B. Mechanism design. Palgrave Macmillan UK, 1989.
4. Is Decision Intelligence The New AI? https://www.forbes.com/sites/forbestechcouncil/2022/05/25/is-decision-intelligence-the-new-ai/?sh=78404df74e42
5. Kraus, Mathias, Stefan Feuerriegel, and Asil Oztekin. "Deep learning in business analytics and operations research: Models, applications and managerial implications." European Journal of Operational Research 281.3 (2020): 628-641.
6. Sutton, Richard S., and Andrew G. Barto. Reinforcement learning: An introduction. MIT press, 2018.
7. Pearl, Judea, and Dana Mackenzie. The book of why: the new science of cause and effect. Basic books, 2018.
END
Author and department introduction
The author of this article, Zhong Chen, is from Didi’s online car-hailing MPT team (Marketplace Technology). The team is committed to building the world's top intelligent trading platform, including order allocation, driver scheduling, carpooling, pricing, subsidies, etc., through continuous exploration of cutting-edge technologies such as machine learning and reinforcement learning, to improve the design of the trading market, to achieve the optimal allocation of resources, and strive to Solve the ongoing and potential imbalance between supply and demand, meet the diversified travel needs of the platform to the greatest extent, continuously optimize passenger experience and guarantee driver income, improve business operation efficiency, and lead the transformation and development of the travel industry.
Job Offers
We are recruiting for the backend of the team and algorithm needs. Interested partners are welcome to join. You can send your resume to [email protected], or scan the QR code below to send your resume directly. Looking forward to your joining!
Senior R&D Engineer
Job Responsibilities:
1. Responsible for the design and development of the core dispatch engine architecture, distributed matching computing system, etc.;
2. Responsible for the architecture design and development of complex strategies such as order distribution, flow diversion, and supply and demand forecasting;
3. Responsible for the exploration of new business models.
Senior Algorithm Engineer
Job Responsibilities:
1. Research algorithms including various transaction matching, order dispatching, and passenger expectations under the modes of solo rides and shared rides, and continue to improve the efficiency of core transactions;
2. Utilize techniques such as causal inference, operational planning, and machine learning to improve the effects of core operational algorithms such as supply and demand forecasting and subsidy pricing;
3. Use algorithm technology to achieve efficient growth of users of each business line of the group and optimize traffic operation efficiency;
4. Solve driver and passenger disputes and experience problems through machine learning technology, create a good driver experience and platform order, build a fair judgment ability for drivers and passengers, and protect the safety of drivers and passengers.
Dry goods learning, like three times ↓