Introduction
The last topic said ensembles need diverse models. One way is to use different algorithms. Another, often more powerful way: use the same algorithm many times, but give each copy a different random sample of the training data. Each model sees a slightly different picture of the world, so each makes slightly different mistakes — and averaging cancels them out.
It's like asking ten students to each study a different random selection of past exam papers, then combining their answers. Each student has gaps, but they're in different places.