Learning rate, epochs, and batch size are the visible knobs, and each has a symptom it produces when wrong. The way to set them is to change one at a time and read the held out score, because two wrong knobs can cancel and leave a number that looks fine.
About 1 min of reading
Before this new lesson, one small question about the lesson you just finished. Recall it, do not look it up.
Compare two runs.
The scores
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Change one knob, rerun, compare the held out number. Changing three at once produces a better number you cannot explain, which is a result you cannot repeat.
Two runs of the same setup can differ. Fixing the seed and recording it is what makes a difference between runs real rather than noise.
Too high and the model overshoots and the curve is jagged. Too low and the curve barely moves. Both symptoms are visible, and both are fixed by moving one notch rather than guessing a new value.
Set by the held out curve, not by a habit. The right count is where held out loss stops falling, and that point moves with the size and variety of the set.
It changes speed and memory. If quality changes with batch size alone, something else in the setup is also wrong, since batch size should not be doing the learning.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.