mathematics//simulation//Monte Carlo method

The Monte Carlo method is a way of computing the distribution of an outcome by running a model many times with its uncertain inputs drawn at random and collecting the results, and engineers use it whenever the question is how widely something can go wrong: how far a drone drifts in gusty wind, how late a fleet's deliveries can be, how soon a bearing might fail. Instead of solving for the spread analytically, it samples it. A thousand simulated landings, each with a wind gust, a sensor bias and a motor constant drawn from their measured ranges, give a thousand touchdown points; their scatter is the answer, and the worst ten of them are usually the interesting part.

The Monte Carlo method is a way of computing the distribution of an outcome by running a model many times with its uncertain inputs drawn at random and collecting the results, and engineers use it whenever the question is how widely something can go wrong: how far a drone drifts in gusty wind, how late a fleet's deliveries can be, how soon a bearing might fail. Instead of solving for the spread analytically, it samples it. A thousand simulated landings, each with a wind gust, a sensor bias and a motor constant drawn from their measured ranges, give a thousand touchdown points; their scatter is the answer, and the worst ten of them are usually the interesting part.

Its accuracy depends only on the number of runs. An estimate built from NNN independent runs has a statistical error that shrinks as 1/N1/\sqrt N1/N​, whatever the model's complexity or number of inputs: four times more runs halve the error. That makes it the method of choice for nonlinear models with many uncertain parameters, where linearized uncertainty propagation is cheap but can be wrong, and it is how a linearized result is checked.

It is the robustness check engineers run most. Simulating hundreds of variants of a vehicle with masses, delays and motor constants spread over their tolerances, and looking at the worst case, is far more common in practice than formal robust control analysis, and it works on any simulator. Its guarantee is empirical: it covers the variations that were sampled.

Rare events are what it handles worst. A failure that happens once in a hundred thousand runs needs millions of runs to estimate with any precision, because the precision is set by the number of failures observed (sample size). Importance sampling and accelerated tests exist for that reason.

Monte Carlo trajectory projection turns an estimate into a forecast. A degradation model whose current state and rate are known with uncertainty is run forward hundreds of times from sampled starting points; each future trajectory crosses the failure threshold at some time, and the histogram of those crossing times is the distribution of the remaining useful life. The same move estimates the queues and the tails of a complex system, where averages mislead.

The name covers several families that share random sampling. A particle filter is sequential Monte Carlo, propagating a cloud of samples through time as data arrive; MCMC draws samples from a posterior distribution that cannot be sampled directly. Plain Monte Carlo, as here, draws independent runs from inputs that can.

A fixed random seed makes a Monte Carlo study reproducible, which is what lets a regression test compare two versions of a controller on exactly the same thousand gusts.