If there is one thing that characterizes driving in any large city, this is a constant stop when the lights change, and cars and trucks merge, separate and rotate and park. This continuous stop and starting is extremely inefficient, increasing the amount of pollution, including greenhouse gases, which are emitted on a mile of driving.
One approach to counteracting is known as ecological lead, which can be installed as a control system in autonomous vehicles to improve their performance.
How much can it do? Would the impact of such systems on reduction of emissions be worth investing in technology? Doing such questions is one of the wide category of optimization problems that were difficult for researchers, and it was difficult to test the solutions they invent. These are problems that affect many different agents, such as many different types of vehicles in the city and various factors affecting their emissions, including speed, weather, road conditions and light time.
“We were interested in a few years ago: is there anything that automated vehicles in the field of emission alleviating can do?” Says Cathy Wu, Thomas D. and Virginia W. Cabot Career Development Professor in the Department of Citizenship and Environmental Engineering, as well as the Institute of Data, Systems and Society (IDSS) in MIT and the main researcher at the Laboratory of Information and Decision Systems. “Is this a decline in a bucket, is there anything to think about?” She wondered.
To answer such a question about so many components, the first requirement is to collect all available system data from many sources. One of them is the system of network topology, says Wu, in this case the map of all intersections in every city. Then there is data from geological tests showing heights to determine the road assessment. There is also data on temperature and humidity, data on a mixture of vehicles and centuries and a mixture of fuel types.
Eco-conducting consists in making small corrections to minimize unnecessary fuel consumption. For example, when the cars are approaching the light that became red: “There is no point in going to the red light as soon as possible,” he says. Just the coast: “I don't smoke gas or electricity in the meantime.” If one car, such as an automated vehicle, releases at the approach to the intersection, then the conventional, uncommon cars behind it will also be forced to release, so the impact of such efficient driving can go far beyond the car itself, which does it.
Wu says that this is the basic idea of ecological driving. But to determine the impact of such measures, “these are problems with optimization” covering many different factors and parameters, “so there is a wave of interest on how to solve control problems using AI.”
The new comparative system, which Wu and its colleagues have developed based on urban eco-conduct, which they call “Ventsectionzoo”, is to help satisfy part of this need. The benchmark has been described in detail paper Presented at an international conference on the representation of learning in Singapore.
Looking at the approaches that were used to solve such complex problems, Wu claims that the important category of methods is multi -stage learning of deep strengthening (DRL), but the lack of appropriate standard reference points for the assessment of the results of such methods hindered the progress in this field.
The new reference point is aimed at solving an important issue that Wu and its team identified two years ago, i.e. in the case of most existing algorithms of learning deep strengthening, when it is trained in the case of one specific situation (e.g. one special intersection), the result does not remain valid when even small modifications were made, such as adding a bicycle path or changing the time of movement, even when they are suitable for light.
In fact, Wu indicates that this unlimited problem “is not unique to move,” he says. “He returns back to canonical tasks, which the community uses to assess the progress in the designing of the algorithm.” But because most such canonical tasks do not involve modifications: “It is difficult to know if your algorithm is making progress in the case of this kind of solidity problem if we do not assess it.”
Although there are many comparative points that are currently used to assess the algorithmic progress in DRL, he says: “This eco-driving problem contains a rich set of features that are important in solving problems in the real world, especially from the point of view of generalization, and that no other reference point.” That is why 1 million traffic scenarios based on data in Intesectionzoo is exceptionally positioned to progress in generalizing DRL. As a result, “this reference point increases the richness of how to assess deep algorithms and progress.”
As for the initial question about urban traffic, one of them will be the use of this newly developed comparative tool to solve a special case of impact on emissions from the implementation of ecological vehicles in automatic vehicles in the city, depending on what percentage of such vehicles is actually implemented.
But Wu adds that “instead of creating something that can implement the ecological driving of a car on a city scale, the main purpose of this study is to support the development of the general grading of deep reinforcement algorithms, problems that can be used to this application.”
Wu adds that “the goal of the project is to provide this as a tool for scientists that is openly available.” VentSectionzoo and documentation on how to use it are free of charge at the address Girub.
The authors of Vindula Jayawardan, a graduate of Mit Department of Electrical Engineering and Computer Science (ECs); Baptiste Freydt, graduate of Eth Zurich; and co -authors of AO QU, a graduate of transport; Cameron Hickert, an IDSS graduate; and Dr. Zhongxia Yan '24.
