In practice we are often in situations where multiple hypothesis explain the data we observe but diverge from each other on data that we have not yet observed and it may be difficult to create the necessary situations to tell between them. In these cases we may still wish to choose a hypothesis to make predictions from.

There are an infinity of possible models to choose from, with most of those models containing no less information than the phenomenon they seek to model. Predictive power is what is important for models; a model that has enough dials to be adjusted to work with any new piece of data might be 'correct' but it is not useful. My favourite example of this is the fact that when the heliocentric model of the solar system was being developed, the geocentric model was providing much more accurate values for the positions of celestial bodies for a long time because it had had years of being tweaked to do so. Initially, it was the simplicity of the heliocentric model rather than its accuracy that was appealing.

If we already know the answer, then we don't need a hypothesis. The only reason to choose is because there are different predictions of the unknown.

"Simple" here has a particular meaning, related to Kolmogorov complexity.

Could you clarify what you mean by that?