Training a model is an expedition across a loss landscape: every point is a setting of the weights, its height is how wrong the model is, and learning means walking downhill.
Gradient descent only ever steps in the steepest-downhill direction, a distance set by the learning rate. Too small and it crawls; too large and each step overshoots the valley floor and lands higher up the far wall — the party ricochets and diverges.
Momentum gives the party inertia: it keeps a fraction of its last step, so it coasts through shallow basins and over low ridges that trap plain descent — though it can also sail past a good minimum. Stochastic noise (the drunk scout) jitters each gradient, letting it rattle out of narrow traps at the cost of a wobbly path.
The lesson: these hyperparameters are personality traits, and no single personality is best. Sculpt the terrain and the same trait that wins on one landscape loses on the next.
The simulation stopped unexpectedly — the lesson continues without it. You can move on; nothing you did was wrong.