Daniel Zuckerman has a website called the "Statistical Biophysics Blog" (http://statisticalbiophysicsblog.org/) that does a great job of explaining some basic principles behind trajectory ensembles. I'd recommend giving that a read, specifically the "More is Better: The trajectory ensemble picture" and "FAQ on Trajectory Ensembles."
The tutorial by Justin Lemkul that Mokhtar Nosrati posted is also helpful in introducing you to the technique.
That said, I'll do my best to briefly explain umbrella sampling with an example from my research. Just for background I study a peptide that crosses membranes. We have run many simulations for long timescales and have failed to see a peptide passively cross a membrane (and there are many possible explanations, like they use active transport or our forcefield is overestimating the partitioning free energy for basic residues, or whatever), but either way, there lies the problem. A long time scale simulation hasn't proven enough to give us the full picture.
So to study the possibility of passive transport, I can first use steered molecular dynamics to literally drag the peptide through the membrane, thus generating a whole bunch of intermediate frames at different depths in the bilayer. Then, I can use umbrella sampling to run a series of simulations (what Zuckerman refers to as a trajectory ensemble) from each of those intermediate frames. In those simulations, I'll restrain the peptide backbone's center of mass to a specific depth.
This will allow me to gain equilibrium information about processes for which there is a large energetic barrier (in this case, the inability for the peptide to partition fully into the hydrophobic core of the membrane).
And yeah, that's basically what umbrella sampling is and what it's for, to study something that's otherwise not readily accessible through a single trajectory.