Hacker News Reader

Comment by kelnos | Hacker News Reader

> There were 9 crewed missions and 1 almost killed its crew (I will omit Apollo 1 for now). I could argue that Apollo had a 1 in 20 chance of killing a crew.

That's not how risk analysis works.

Let's say every Apollo mission had gone flawlessly, and no one even came close to dying. Would you then say that the risk of death for future missions would be zero? No, of course not.

View on HN

I thought I'd look this up. If you've had 9 successful attempts, assuming nothing has changed between them and no other prior knowledge about success probability, then Laplace’s Rule of Succession says the probability of the next mission being a success is about 83.3%, i.e. there is a 1 in 6 chance of failing next time.

> and no other prior knowledge about success probability

This phrase is misleading, as Laplace's Rule of Succession is equivalent to assuming a uniform Bayesian prior over all values of p. That is, before any experiments, a 50% chance of success. Depending on the situation, this may be roughly accurate or wildly wrong. You cannot appeal to this rule to resolve the situation.

Well, obviously if we have a better prior, then that's better. But assuming no other knowledge, and especially if we think that other people's priors could be intentionally misleading, this rule seems to offer the best estimate.

Generally speaking, you never truly have "no prior knowledge". Some relevant past experience, or "common sense", or something tips you away from "all probabilities are equally likely". I think this rule is rarely a best estimate.

If you only look at this data it would be the most reasonable guess.

Of course one should also analyze the technical sytems involved and then it is clear that 0% failure is not reasonable.

Bayes rule has existed for nearly 300 years, there is no excuse for ‘only look[ing] at this data’ and that is NEVER a reasonable thing to do.

Reply on news.ycombinator.com