Does the A* algo start from the desired goal, or from the current world state?

[ghost]

When I was reading about GOAP few weeks (or was it months?) ago, I could find that the F.E.A.R. original implementation started from the desired goal instead of the current world model.

The argument, which made sense to me (but I'm rather new to the topic of game bots or automatons, and don't have bagage on the area), is that it could invalid the plan sooner if some actions are impossible.
This being true or false (or me misunderstanding it) won't make it least interesting or useful to document it though, (helps trying to understand bugs in our codebases :)), and if possible the reasons for the choice would be nice (simpler to implement being perfectly valid).