Safe Systems from Unreliable Parts

My boss and I found a nice example of this in some sampler-placement work we did. The punchline is that our software occasionally doubled-up samplers in order to increase the probability of detecting a release.

The problem domain is a biological or chemical release in a building. Given a suite of possible release scenarios (variable release locations, amounts, and dispersion conditions), find the sampler network with the highest probability of detecting a release. Pretty much everything is probabilistically defined, including how likely each scenario, and how likely a sampler is to detect a given mass of the released substance. Critical to the discussion here, this probability was assumed independent of what other samplers did.

The software just brute-forces its way through all possible sampler networks. One striking result from our (very simple) study case was that it often put two samplers in the same location. This is a clear consequence of the principles laid out in the linked article.

For further info: Sohn and Lorenzetti, Risk Analysis, v.27 n.4 (2007), pp.877-886.

When you do this with mortages you get a financial crisis!