Hacker News Reader

Comment by smartician | Hacker News Reader

Aren't many AI methods also inherently probabilistic? That would make it by definition impossible to definitively explain any particular behavior.

View on HN

If you have a probability vector, you would typically take the max, not model the action on the probability distribution.

e.g. If an image (from the forward camera) returns the probability vector: {"stop_sign": .9, "green_light": .1} You stop the car 10/10 times.

What do you mean by probabilistic? If you mean that the models takes input and outputs one or more labeled probabilities (e.g. “90% confidence the input photo is a dog, 10% confidence the input photo is a cat), then yes, I believe that many AI systems work this way. If you mean random, in the sense that the system may return different results given the exact same inputs, then in not sure if there are AI systems that work that way.

There are. Monte Carlo methods is the keyword you're looking for. AlphaGo (Monte Carlo Tree Search) is an example of one such AI.

Obviously you can set the RNG seed to be the same every time too, but even that only works if your system is wholly synchronous, which a car probably isn't.

Note that I doubt Monte Carlo methods are common in the autonomous vehicle space.

Well I don’t work in autonomous cars but I did work on several autonomous robots, and “ Adaptive Monte Carlo Localization” is a common way to keep a ROS + LIDAR based robot localized.

I wouldn’t be surprised if Monte Carlo techniques are useful for all manner of things related to ingesting sensor data on autonomous vehicles.

I'd expect Monte Carlo methods to be used in a number of cases that have deterministic time envelopes for evaluation. Randomized selection and evaluation can be incredibly effective. They also resist degenerate structured input vulnerabilities.

I'm not in the automotive space, but I'd be surprised if there were a viable self-driving car team not using Monte Carlo methods somewhere in the vehicle stack.

Yeah, this was unfortunately worded. There are subproblems for which MC/randomized methods fit well, but in general those circumstances are well understood.

I don't think that means it's impossible. You can still log all inputs and sources of entropy.

Reply on news.ycombinator.com