Space Apps Challenge 2017

The project's goal is to demonstrate how artificial intelligence and machine learning algorithms supported by GNSS data and other data (e.g., cameras, inductive loops, road monitoring by drones / unmanned aerial vehicles) may help improve tranport and mobility. We used and developed further a code of TensorTraffic tool (under development since November 2016) https://github.com/pgora/TensorTraffic to train neural networks to approximate some traffic simulation outcomes. In our experiments we were approximating total waiting times on a red signal in a small area of Warsaw ("Stara Ochota" district,15 crossroads with traffic signals) and used it to find (sub)optimal traffic signal settings in that area, using a genetic algorithm. It gave us opportunity to find traffic signal setting reducing the total waiting times on a red signal by more than 28% comparing to the best setting in a large initial (random) set of configurations. Our genetic algorithm was able to find such good setting in 44 seconds (we expect that this time can be even reduced by using GPU). We proved in experiments that our approach for traffic simulation approximation may be better than linear regression and random forests, giving a very low average error (~1.2%). Also, thanks to using GPU it may be possible to obtain results of neural networks inference a few orders of magnitude faster than in case of standard traffic simulations. To run experiments we used standard virtual machines with GPU K80 in the Windows Azure cloud computing infrastructure.

Our approach may be potentially used for real-time traffic management (e.g., traffic signal control) or in any other situations requiring evaluating large number of possibilities, e.g., to desing optimal road infrastructure (e.g., location of bridge which may minimize time spent by people in travel or distance to important buildings), finding optimal locations and capacities of parkings and charging stations for electric vehicles etc.

This approach may be especially successful in the era of connected and autonomous vehicles, because their routes and trajectories of drive can be much better predicted than in case of human-drivers.

The similar approach may be also used for traffic management of drones and unmanned aerial vehicles when their number in the air will be so large, that there will have to exist aerial roads and traffic management centers ensuring traffic safety and efficiency.