This kind of complex-system-optimization problem seems ripe for improvement via machine-learning. Cases where 1) precisely modeling the outcome is totally intractable, and 2) you really just need to get "close enough" to get good results. It also helps that this case is predictive in nature (so it's easy to go back and measure accuracy) and has an enormous set of available training data (all traffic happening everywhere across the country).
It feels very similar in spirit to that story about Google reducing data center energy use by 40% via ML optimization: https://deepmind.com/blog/article/deepmind-ai-reduces-google...
More broadly I've started wondering if there's a huge opportunity to apply software to civil improvement. Everything from detecting patterns of corruption, to aggregating data that can inform voters, to identifying sources of inefficient spending. I don't put much stock in things like digital voting/governance, but in the area of "data crunching and presentation" I think there's a lot that could be done.