Even if I do not have time, I simply had to write this article. As I said, there is a biological limit to how much you can actively work in a week. In addition to that, one can still do passive things like reading or watching Youtube videos. The algorithm has now suggested a very interesting video for me about neural networks:
The whole thing is related to a project, which I did not mention in my previous article because I did not pursue it. But only because of the time / benefit relation, the topic is extremely interesting.
This story begins in my childhood days while on the road to vacation: how many times have I wondered if you could not program a simulator that explains why you're stuck in a traffic jam right now. That was the time of the C64 computers back then and I had no idea about programming, but the idea was there. I revisited the topic after the 2016 CityCampus in Trier: I asked one of the project teams for applied computer science (who were concerned with the prediction of traffic flows), why nobody would be work on the question of how to program really intelligent traffic lights, but got no conclusive answer. You have to know that there is sometimes a kilometer-long traffic jam in Trier on the main roads and sometimes everything is free. The responsible traffic light stubbornly pulls off its timed program. Anyway, I then started programming a traffic simulator for two full days, where individual cars with random properties would drive one-dimensionally at first. The result also looked pretty nice, apart from the occasional rear-end collision (the braking had probably a bug in it). I then had the idea that a road and traffic network is basically just a big system of equations, that you have to optimize for the total waiting time. The how was not defined yet, I thought that you copy the reality into a simulator and then brute force optimize the ideal traffic light phases for this point in time.
Now, when I saw the video about the neural networks, I realized that the mechanics of learning the network due to the gradients is exactly the optimization I would have needed for this system. A system of equations with hundreds or thousands of variables can not be solved otherwise, since there is no compelling, distinct solution. The network is likely to have a different structure than a multi-level abstraction network, but the underlying mechanism is the same. A simulation with the control variables is made, cost is the accumulated waiting time of the simulated vehicles. The special thing is the interaction of the neural network for control with the simulation, which is why there are no linear relationships between the nodes. With stochastic optimization you would still have to go further, there is still a need for development. All in all, this is an even more intriguing idea, with a tiny little catch: when should I do that and does it evolve into anything tangible? All of my other projects have a tangible benefit, here the likelihood is extremely low (and by far exceeds my resources) that it will become a functioning system of urban transport planning. If there was not the curiosity ... and worse, I'm pretty sure I could make such a system, given enough time.
I have already even thought about the implementation: you have to make the city as intelligent as possible, so sensors and cameras capture the cars (and maybe pedestrians) in as many places, everything then additionally with the temporal component. Then you know exactly that at the intersection X on a Wednesday at Y clock Z vehicles wil arrive and 23% of the vehicles turn left and so on. This information is fed into the simulation and allows the neural network to learn the ideal control, because in the simulation, the network can play around without causing damage. Even if you do not have 100% coverage at first, that's not bad, you can supply individual nodes with estimated values, because of that the simulation does not collapse. When the optimal solution is found, it is put back into real life. Even if there is only a few percent optimization multiplied by the number of people involved, there must be a huge gain of lifetime.