In many practical situations, we cannot directly measure or estimate the desired quantity y – e.g., we cannot directly measure the distance to a star or the next week’s temperature. To provide the desired estimate, we measure or estimate easier-to-measure quantities x1, … , xn which are related to y, and then use the known relation to transform our estimates for xi into an estimate for y. In situations when xi are known with fuzzy uncertainty, we thus need fuzzy computation. Zadeh’s extension principle provides us with formulas for fuzzy computation. The challenge is that the resulting computational problem is NP-hard – which means that, unless P=NP (which most computer scientists consider to be impossible), it is not possible to solve all fuzzy computation problems in feasible time. To overcome this challenge, we propose a more realistic formalization of fuzzy computation – in which instead of an un-realistic requirement that the corresponding properties hold for all xi, we only require that they hold for almost all xi – in some reasonable sense. We show that under this modification, the problem of fuzzy computation becomes computationally feasible.