5 ms·

Yes, but in this case compliance is quite difficult. Why would the compliance department go out of its way to develop a new mobile testing infrastructure when t

Yes, but in this case compliance is quite difficult. Why would the compliance department go out of its way to develop a new mobile testing infrastructure when that isn't the industry standard? So it seems very plausible in this case that compliance might not have found the issue. It required a novel method of testing to discover the cheat. I have to be honest, when I first heard about this, I suspected a genetic algorithm as the culprit. It's exactly the kind of result you'd expect when giving an impossible optimization without sufficient boundary constraint. AIs can and have been known to cheat, and because the fitness driver would have been compliance (V&V) the GA could plausibly have settled on methods that were undetectable (i.e. Not part of the fitness function). Often the output of GA code is not understood formally by engineers, or reverse-engineered into a traditional algorithm, so it's plausible that it could have remained hidden in plain sight. BUT, if that had truly been the case, the engineers at VW would have quickly suspected and publicized it. So the fact that VW blinked showed that in fact they lied and didn't have an honest technical explanation. However, I think the GA scenario is good to keep aware of as we increasingly rely on machines to optimize solutions based on testing.

jackgavigan 11y ago
> Yes, but in this case compliance is quite difficult. Agreed. It's highly unlikely that the Compliance department could have detected what the engineers had done, short of having someone audit the code, line by line, or inventing a test to check the emissions while the car was moving. Based on my experience in the financial sector, Compliance departments don't work like that - if someone really wants to cheat or circumvent the rules, there's a good chance they'll be able to conceal it from Compliance, at least for a while. People are responsible for their own actions, even when their job title is "engineer".
- illumen 11y ago
  VW modeled specifications and tested them for other components. Why wouldn't these components be written to specification?
- jacquesm 11y ago
  We're not talking 1.1 vs 1.2, we're talking a 40 fold increase. If the compliance department is not able to detect such a massive fraud then they should all be out of jobs. And if they don't audit the code then they should be out of jobs anyway. There is a big difference between process compliance (the kind that the financial world is burdened with) and product compliance. Independent laboratories would have been able - and in fact did - to find these discrepancies with comparable ease, further down there is a link to a magazine article where 9 cars were tested by an outside lab just to write an article.
  - jackgavigan 11y ago
    The magnitude of the increase is irrelevant. The key question is whether it was reasonable to expect them to detect it, which is, in turn, informed by the question of how they could have done so. The obvious method to audit compliance against an emissions standard is to use the same test that the regulator does. Obviously, in this case, that wouldn't have detected the defeat device. To detect it, they would have needed to either conduct a non-standard test, or audit the code. Both those solutions seem, to me, to be above and beyond what would be considered reasonable due diligence under the circumstances. Perhaps I'm wrong, though - perhaps other car-makers perform such tests as a matter of course? I'm happy to be presented with evidence that that's the case. Even if it is the case, however, that still doesn't absolve the engineers from blame. The bottom line is this: You are suggesting that the engineers who created the defeat device should not be blamed for doing so because the Compliance department should have detected it. I'm saying that not only was it not necessarily reasonable to expect the Compliance department to detect it but, even that were the case, it doesn't absolve the engineers of responsibility for what they did.
    - jacquesm 11y ago
      I suspect the engineers were ordered to do this, not that they came up with this by themselves. (In fact, I find that suggestion totally unbelievable.) I agree the engineers are culpable but if they were acting under orders from their managers then the buck has to stop somewhere, someone is going to be the last one 'up the chain' that decided it was within their pay-grade to make this decision. The internal organization of VW is such that that decision could have only come straight from the top given the liability in case of exposure. If the only way to ensure compliance is to audit the code then that's what you should do. It basically says the testing regime is so easy to detect that it can be used to defeat the test itself, Bosch correspondence indicates that this software is expressly not meant to be included in the final product and VW knowingly ignored that information, something a compliance department should have been absolutely on top of.
ArkyBeagle 11y ago
The "sin" they are accused of it explicitly detecting that certain test equipment is connected. That's a binary, if-then-else sort of thing. "The boss is looking; look clean." A GA (or any machine learning ) within a controller is less about being able to cheat than it is ...rebasing something that would otherwise be a constant (from the frame of reference of the algorithm) in a control-feedback arrangement. If the GA was part of a test stand and exhibited emergent... dishonest behavior :) then ... they did it wrong.
- Shivetya 11y ago
  There are valid reasons to write code to determine if your vehicle is being tested. Modern cars contain all sorts of computerized safety features that would go active in the typical test scenario thereby preventing the static tests that the EPA and others do. So the car detects wheels straight, only drive wheels turning, and so on, so it disengages the features that would active traction control, reducing engine power, and more. The key here is that VW engineers went further and used that as they indicator to run the full emissions process and only then. There is no viable reason to have any toggle on emissions processes except perhaps with regards to cold weather starts.
  - ArkyBeagle 11y ago
    We think it's "run the full emissions process" but we're guessing. It's a very good guess, but it's essentially unsupported. It could have been an interaction with one of the other things that got turned of. This seems unlikely,. but it's not 0% probability.
- coldnebo 11y ago
  Doing GAs right can be very hard. Like, I've seen cases where GAs hooked up to a certain type of physical simulation learn to exploit the ODE in play and find optima that are based on ODE artifacts, not reality. Eg wind turbine design and CFD modeling. I remember one specifically learned in Quake to always strafe run because the dx+dy gave sqrt(2) max speed instead of just max speed. What most people don't understand about GAs is that asking the question correctly and unambiguously can be harder than solving the problem. GAs can have a steep price, it's not magic.

jackgavigan 11y ago

> Yes, but in this case compliance is quite difficult. Agreed. It's highly unlikely that the Compliance department could have detected what the engineers had done, short of having someone audit the code, line by line, or inventing a test to check the emissions while the car was moving. Based on my experience in the financial sector, Compliance departments don't work like that - if someone really wants to cheat or circumvent the rules, there's a good chance they'll be able to conceal it from Compliance, at least for a while. People are responsible for their own actions, even when their job title is "engineer".

illumen 11y ago
VW modeled specifications and tested them for other components. Why wouldn't these components be written to specification?
jacquesm 11y ago
We're not talking 1.1 vs 1.2, we're talking a 40 fold increase. If the compliance department is not able to detect such a massive fraud then they should all be out of jobs. And if they don't audit the code then they should be out of jobs anyway. There is a big difference between process compliance (the kind that the financial world is burdened with) and product compliance. Independent laboratories would have been able - and in fact did - to find these discrepancies with comparable ease, further down there is a link to a magazine article where 9 cars were tested by an outside lab just to write an article.
- jackgavigan 11y ago
  The magnitude of the increase is irrelevant. The key question is whether it was reasonable to expect them to detect it, which is, in turn, informed by the question of how they could have done so. The obvious method to audit compliance against an emissions standard is to use the same test that the regulator does. Obviously, in this case, that wouldn't have detected the defeat device. To detect it, they would have needed to either conduct a non-standard test, or audit the code. Both those solutions seem, to me, to be above and beyond what would be considered reasonable due diligence under the circumstances. Perhaps I'm wrong, though - perhaps other car-makers perform such tests as a matter of course? I'm happy to be presented with evidence that that's the case. Even if it is the case, however, that still doesn't absolve the engineers from blame. The bottom line is this: You are suggesting that the engineers who created the defeat device should not be blamed for doing so because the Compliance department should have detected it. I'm saying that not only was it not necessarily reasonable to expect the Compliance department to detect it but, even that were the case, it doesn't absolve the engineers of responsibility for what they did.
  - jacquesm 11y ago
    I suspect the engineers were ordered to do this, not that they came up with this by themselves. (In fact, I find that suggestion totally unbelievable.) I agree the engineers are culpable but if they were acting under orders from their managers then the buck has to stop somewhere, someone is going to be the last one 'up the chain' that decided it was within their pay-grade to make this decision. The internal organization of VW is such that that decision could have only come straight from the top given the liability in case of exposure. If the only way to ensure compliance is to audit the code then that's what you should do. It basically says the testing regime is so easy to detect that it can be used to defeat the test itself, Bosch correspondence indicates that this software is expressly not meant to be included in the final product and VW knowingly ignored that information, something a compliance department should have been absolutely on top of.

ArkyBeagle 11y ago

The "sin" they are accused of it explicitly detecting that certain test equipment is connected. That's a binary, if-then-else sort of thing. "The boss is looking; look clean." A GA (or any machine learning ) within a controller is less about being able to cheat than it is ...rebasing something that would otherwise be a constant (from the frame of reference of the algorithm) in a control-feedback arrangement. If the GA was part of a test stand and exhibited emergent... dishonest behavior :) then ... they did it wrong.

Shivetya 11y ago
There are valid reasons to write code to determine if your vehicle is being tested. Modern cars contain all sorts of computerized safety features that would go active in the typical test scenario thereby preventing the static tests that the EPA and others do. So the car detects wheels straight, only drive wheels turning, and so on, so it disengages the features that would active traction control, reducing engine power, and more. The key here is that VW engineers went further and used that as they indicator to run the full emissions process and only then. There is no viable reason to have any toggle on emissions processes except perhaps with regards to cold weather starts.
- ArkyBeagle 11y ago
  We think it's "run the full emissions process" but we're guessing. It's a very good guess, but it's essentially unsupported. It could have been an interaction with one of the other things that got turned of. This seems unlikely,. but it's not 0% probability.
coldnebo 11y ago
Doing GAs right can be very hard. Like, I've seen cases where GAs hooked up to a certain type of physical simulation learn to exploit the ODE in play and find optima that are based on ODE artifacts, not reality. Eg wind turbine design and CFD modeling. I remember one specifically learned in Quake to always strafe run because the dx+dy gave sqrt(2) max speed instead of just max speed. What most people don't understand about GAs is that asking the question correctly and unambiguously can be harder than solving the problem. GAs can have a steep price, it's not magic.