I had a thought on my walk that maybe uncertainty analysis and risk analysis is a useful area in sustainability. Maybe it occupies that ‘underappreciate’, ‘scalable’ thing EA talks about. Something I read that micheal nielsen said also was that often people don’t focus on gaining expertise in some X when Y pops up. That often to get really good at that X rather than jump to the new Y is an efficient thing to do.
The thing is, I’m not any expert at this stuff, nor would I say I’m particularly skilled. In saying that though it is underrated and I think its useful. At least, it might keep us from fooling ourselves by forcing us to formalise our assumptions and work through what they actually mean for some target.
Research
If theres anything to the Helton paper I think this is it. I need to also get more familiar at categorising analytical uncertainty. It seems basic in my head but I can’t quite communicate it.
This is kind of in the hope that in the midst of the linguistic definitions of uncertainty there is some notion of formalism.
The language of uncertainty is probability. The most general form of uncertainty analysis is a determination of the distribution of an output of a function for some set of input distributions (helton). The effect of this uncertainty on model output depends on the funciton mapping input to output. It is called uncertainty propagation and there are several techniques to perform it. When there are multiple paramters the paramter domain space becomes intractible. Integration over this space becomes computationally expensive. Numerical methods are used to approximate the integral. Monte Carlo sampling involves
This general form can break down the sensitivity analysis too
Sensitivity analysis can then be viewed as a decomposition of V(y) into components due to the individual elements of x, with the size of these components then providing an indication of variable importance.
When to use MC sampling
Approximation of a function (the output probability distribution) where there is only one input and the function is linear does not require sampling. The integral is most likely calculable.
Where it is needed is when the amount of inputs increases. If we think of the output as a path function where the path is some selection of random variables corresponding to the inputs domain random variables. Here we would have an n-dimensional integral to calculate a fully defined distribution for the output. We approach this function with numerical methods like MC sampling.
Uncertainty analysis of feedstock transport
- distances for shipping. Routes it chooses. Could have different routes, as opposed to a road.
- Storage time in tank.
- Emission factors for truck.
- Liams fuel yield affects allocation.
- Sensitivity analysis for each input
- Sampling for output distribution .