Machine learning techniques are already being used for autonomous cars that are being developed, whereby these cars can handle road conditions under different circumstance and where sensors are mostly within the car. In these vehicles, responses are made from locally gathered data. However, this does not necessarily solve a persistent problem on many streets in the US and other countries. That being traffic.
Traffic is a frequent problem that not only causes consternation among drivers but it contributes to pollution and money and time wasted. One way that traffic could be limited is simply by optimizing speed in which highway and other traffic could move, where the maximum speed limit is set by the best speed for given real-time traffic conditions. This can be done by reinforced learning techniques that simulate the best speed at a given instant. Another way to improve traffic is to integrate technologies within autonomous cars, particularly as they are forecast to be soon on our streets, including sensors for driving, and information from different external sources, such as traffic information from cellphone, and other sources, including satellite data. This idea, created in a project called CIRCLES (Congestion Impact Reduction via CAV-in-the-loop Lagrangian Energy Smoothing) led at the Lawrence Berkeley National Laboratory, is to take this information and use deep reinforced learning to then optimize route selection and driving behaviour by autonomous vehicles so as to minimize traffic build-up. The idea is if cars move smoothly then traffic will less likely build, as a lot of traffic is often caused by uncoordinated and unoptimized movement choices. On the other hand, autonomous vehicles could coordinate their actions with relevant information and optimize their choices through reinforced learning. Related to this, the Berkeley effort will take satellite imagery and combine it with cellphone data to determine pollution levels, where cellphone data provides information on the number of road users and satellite data on pollution levels. Convolution neural networks or similar approaches can be used to identify relationships between pollution and road users that could then be fed into models that forecast pollution levels, where the results could then be fed back into traffic optimization. The pollution project is called DeepAir.
The Berkeley project is based on Flow, which is an open source platform that is used to simulate and optimize traffic flow. Similarly, other projects have also tried to use fast simulation that look at current conditions, simulate those conditions in the immediate future, and then those results are fed back to the vehicle for information to optimize driving. This approach uses neural network-based learning, similar to Flow, to enable fast and rapid learning to better optimize movement.
This side-by-side simulation from Cathy Wu at UC Berkeley shows how one autonomous car in a traffic loop of 22 cars (with the rest human-driven) can smooth out the traffic flow by reducing stop-and-go traffic:
There are other potential variables that are also relevant, including how autonomous cars can move. For instance, optimizing how road traffic signals communicate with vehicles may require further coordination between vehicles, other data sources such as cellular data showing traffic conditions, and traffic lights. In this case, it has been shown that genetic algorithms, utilising mutation genes in the approach, could be applied to optimize traffic movement, requiring relatively few fitness functions evaluations needed.
Autonomous cars are arriving but how will they improve traffic? New technologies are developing that enable cars to obtain information from multiple sources, including satellite data, and utilise simulations that optimize learning algorithms that then inform cars the best routes to select and at what given speeds based on apparent conditions. These new changes promise to make our lives easier but more importantly have a large effect in reducing pollution, where vehicles constitute about 30% of the pollution in the United States.
Dan Work of Vanderbilt University conducted research to demonstrate how introducing a single autonomous vehicle into a loop of human-driven cars can prevent stop-and-go traffic waves and create a smooth traffic flow.
 For more on speed controls using optimization, see: Walraven, E., Spaan, M. T. J., & Bakker, B. (2016). Traffic flow optimization: A reinforcement learning approach. Engineering Applications of Artificial Intelligence, 52, 203–212. https://doi.org/10.1016/j.engappai.2016.01.001
 For recent information about the project, see: https://newscenter.lbl.gov/2018/10/28/machine-learning-to-help-optimize-traffic-and-reduce-pollution/.
 For more on neural network optimization using fast simulations of real-time traffic, see: Gora, P. (2018). Simulation-Based Traffic Management System for Connected and Autonomous Vehicles. In G. Meyer & S. Beiker (Eds.), Road Vehicle Automation 4 (pp. 257–266). Cham: Springer International Publishing. https://doi.org/10.1007/978-3-319-60934-8_21.
 For more on optimization using genetic algorithms for road traffic controls, see: Jin, J., Ma, X., & Kosonen, I. (2017). A stochastic optimization framework for road traffic controls based on evolutionary algorithms and traffic simulation. Advances in Engineering Software, 114, 348–360. https://doi.org/10.1016/j.advengsoft.2017.08.005.
A recently published book features research on using Google Earth Engine to access and analyze GIS Data.
EO College, a platform for accessing MOOC (massive open online courses) about remote sensing along with other related materials was…
Small scale satellites are changing how we can acquire our data as geospatial analysts.