Hacker News Reader

Comment by torpfactory | Hacker News Reader

In my experience as a mechanical designer physical testing, especially system level integrated testing, is the only way to find errors in the “unknown unknown” category. In a complex system (mechanical, electrical, software, biological, basically all of them), there usually exist interactions that are very difficult to predict. Software-based simulation can help, but is only as good as the models which are used to describe the underlying physics. In my experience, these models aren’t nearly good enough, and without testing, you can’t even say what is wrong with them.

Full-up testing isn’t always possible, but it shouldn’t be abandoned only because it is costly or difficult.

View on HN

One of the things I continue to find disturbing when interacting with the markets we work in (e.g. self-driving & other highly-automated systems) is the growing hope that simulation will somehow bail people out of having to do real world validation, let them take significant shortcuts, or make assurance claims that can't really be made.

We deal in formal verification for some aspects of what we do, and that lead to a conversation that went roughly like this:

AV Exec: "Can we use formal methods to prove that our simulator is as good as reality?"

Me: "No. You can use formal methods to prove your simulator implementation more closely adheres to your simulation model."

AV Exec: "Isn't that the same thing?"

Me: "No. Your simulation model is definitely wrong."

AV Exec: (looking at me both disappointed and despondent)

Me: "Look. It's important to understand that simulation isn't creating an approximation with a known correspondence to reality. Simulation is fabricating an entirely new reality with an opaque correlation to our reality."

AV Exec: "Then how am I supposed to use simulation to prove our system is safe?"

Me: "You can't."

Despair on both sides of the table ensues.

I'm lucky that I'm generally met with people who trust me when I say the simulation is not the reality.

I agree with what your saying. As they optimise more and more, they "fly closer to the sun". That would be ok as long as the models/sim tools were accurate, but as you say they are not so much, which means there is not as much room for error.

Reply on news.ycombinator.com