/PRCWT/ Shenzhen, September 15——On September 8,2023 Tencent Global Digital Ecology Conference Smart Transportation was held in Shenzhen. Liu Zhiyuan, professor, doctoral supervisor and deputy dean of the School of Transportation of Southeast University, delivered a keynote speech entitled "Technical System and Application Scheme of Multi-mode Transportation Large Model".
With the development and application of large models in the artificial intelligence industry, practitioners in the transportation industry are also very concerned about the application scenarios of large models in the industry.
Liu Zhiyuan believes that the transportation industry should develop a transportation large model system that is designed and optimized for the transportation field based on the general large model. This system inherits the two functions of answering questions and creating ability of the general large model, and at the same time realizes the significant improvement of the accuracy of the traditional traffic model on the basis of the three key elements of computing power, algorithm and data.
Liu Zhiyuan said that the traffic model is different from the general model that focuses on the text and visual fields, and should focus on various participation factors in the traffic field, focusing on various problems and challenges of single points, trunk lines and road networks.
Liu Zhiyuan pointed out that the functional barriers between various management departments in the transportation industry are often the bottleneck of transportation applications, and the large model can break through such barriers, provide integrated solutions for various departments, and build a unified underlying foundation for transportation applications, so as to achieve the goal of reducing costs and improving efficiency and improving transportation efficiency.
However, Liu Zhiyuan also stressed that the current traffic model is still in the early stage of development, and in the future, with the continuous improvement of computing power, algorithms and data, as well as the increasing number of traffic applications, the traffic model will also help the transportation industry to better achieve the goal of intelligence.
The following is the full text of Liu Zhiyuan's speech:
Thank you very much for the invitation of Tencent, it is a great honor to exchange our experience and experience in large models at this event.
Under the background of the digital wave, the large model is the core keyword of the industry this year. We believe that the transportation industry should develop its own transportation large model system on the basis of the general large model, which must first understand the background of the large model.
The reason why large models can bring a new paradigm and technological revolution comes from the three key elements of computing power, algorithm and data. The focus of the Grand model is on the generative concept, which refers to the various conclusions and knowledge that are not produced in the productive life of the human knowledge system. This is possible because there is a very large knowledge base behind the large model. Therefore, when we build our own technical system, we need to think deeply about the troika of computing power, algorithms, and data, and master the core technologies inside.
Back to the traffic engineering itself, our current common traffic model is reflected in different links such as construction management, maintenance and transportation, planning and design, and is a traditional system represented by the four-stage and three-parameter model. It has a very deep theoretical support, but its biggest problem is the lack of precision. Taking the three-parameter model as an example, its objective fitting error can reach 26% under the premise of obtaining abundant data. Taking the four-stage model as an example, the error can be as high as 150% in practice. Therefore, such a model system is more a product of speculation, rather than a system built under the objective environment where the process can be repeated and the conclusion can be falsified.
Today, if we want to really use computing power, algorithms, and data in the deep architecture of traffic engineering system, we need to construct our own large traffic engineering model like ChatGPT, which we call multi-modal traffic large model. It can solve the problems of unreasonable assumptions of the existing model and fewer model parameters, and the most important goal is to improve the accuracy and bring great intelligence improvement.
ChatGPT has two core functions, the first is to answer questions, and the other is that it has the ability to create. The same goes for large models of traffic, which can answer my questions about how to control highways, for example, or devise new solutions to existing challenges.
Based on this background, we designed the overall technical architecture of the large model. The training process of large models takes a lot of computing resources and time. After the model is trained, scheme evaluation is the part of answering questions, which is equivalent to the concept of forward calculation. Solution generation, that is, this part of the optimization problem, is equivalent to the concept of backward calculation. Under such a technical system, we clearly put forward our technical system, which is a large model system of large-mode transportation that integrates various data and various elements. This large model system should be described from three different angles: point, trunk and network.
Let's start with a la carte. In short, we can think about how to innovate models and methods based on large amounts of data when there is only one detector. Under the single-point large model system, a very important breakthrough is that the people, cars, roads, and environment faced by a single point should all be the parameters of the model, and the four elements of traffic should be the objects that the model can control and optimize. The current single-point traffic model only takes people and vehicles, that is, demand ontology, as variables, but the system of the large model should take all these elements as variables and parameters, and finally the model itself can calculate the optimal traffic control and signal timing scheme. This is fundamentally different from the previous traffic system, and it is also the accuracy that the computing power, algorithms and data of the large model can bring to us, and finally realize the function of actual traffic engineering.
In the construction of models and algorithms in this respect, we use Gaussian process as a system to build a single point large model of traffic. It is very important to use Gaussian process because its variables and parameters are extensible. When we face the actual complex problems, it should be expanded from the original three parameters to 1000 parameters, 10,000 parameters, can have a very elastic scalability.
Another point is the trunk line. If there are two detectors on the highway, there is an undetectable range between the two detectors. Then through the large model, in the case of limited detector coverage, the model itself can be used to deeply characterize the trajectory and process of the entire traffic flow.
A very important model system is embodied in deep reinforcement learning. Deep reinforcement learning now tends to analyze only a limited number of agents, no more than 1,000. But in the traffic problem, a city like Shenzhen may be tens of millions of agents. So earlier this year, we published a new system called Integrated reinforcement Learning (RL), which addresses the problem of how large-scale deep reinforcement learning in applications can be organically integrated with extremely complex and random traffic systems.
The third point is how to analyze under the road network. A very important feature of the road network is that its detectors are more limited, and most of the road sections are blank. To solve this problem, our proposed solution is to build a transfer learning model based on the information obtained from the existing detector sections, and then migrate to every inch of the road network. Therefore, we do not need to lay large-scale detectors, and the current computing power, algorithm, data and the ability of the model itself can fully achieve the full coverage of the entire network perception level.
Under such a system, based on the grand model construction of trunk line, single point and network, we create a new paradigm of traffic grand model. The key of our technical system is that different physical problems and physical objects involved in traffic engineering should be treated differently. So when we put such a model system on Tencent Cloud, the external interface is a unified interface like ChatGPT, but below the surface is the integration of multi-task learning of single points, trunk lines and road networks.
Under such a technical system, look at the specific traffic applications. A very important feature in traffic applications is the Kowloon water control, and the data barriers and functional barriers between different departments are often the bottleneck of practical problems. Can we make the big model based on the data of different systems, and finally provide an integrated solution, so that the big model brings specific functional support to each department?
We finally give a system solution, and the first answer sheet is still placed on the smart highway, including the smart highway. The system of smart highway is the most complete, so based on the six functions of smart highway construction, management, maintenance, transportation, planning and design, we propose ten different systems that the large model can play a role in. Among the ten systems, based on the accurate characterization of the movement law of people flow, traffic flow and logistics by the large model we built, the original only 3 types of parameters have been expanded to 100 types and hundreds of millions of parameters. Such a huge model system and the support of technical tools have brought paradigm revolution to the entire application system.
There are many examples of how such a paradigm can work in specific planning and design tasks. First, a very important application of the single point large model is in the maintenance of highways. The most classic case of smart highway and machine vision is to use the machine vision of inspection vehicles to do automatic replacement of road maintenance. But at present, these inspection vehicles are often single-unit combat, running slowly, and the recognition accuracy of various complex scenes is also poor. However, if we integrate all kinds of large traffic model tools and inspection vehicles, we can make the inspection vehicles four or five times the speed of the current detection, reducing cost and increasing efficiency while bringing a huge improvement in accuracy. The same goes for the trunk model. It can take all the elements of people, cars, roads and environment as part of the model, and output the optimal control scheme with high precision after calculation. At the road network level, when the existing traffic plan is fine-tuned, it is often based on norms and standards to exert human initiative, which brings a lot of randomness and uncertainty. After finding abnormal conditions of traffic flow on urban sections and expressways, the large model can fine-tune the traffic design based on the ground-to-air cruise of drones under the guidance of traffic flow conditions.
The large traffic model is like a newborn baby, and it continues to grow under the three powerful driving factors of computing power, algorithm and data. Basic theory is its foundation, and application is its vitality. Under the system of large models, we can finally achieve intelligent improvement through integrated model construction. Thank you.