MBGD performs about 156 updates per epoch, each using 32 houses, balancing efficiency (via vectorization) and stability (less noisy than SGD). 

MBGD does not require a learning rate hyperparameter, since averaging over mini-batches eliminates the need for step size control.

MBGD performs only 1 update per epoch, using all 5,000 houses at once, giving deterministic but very slow convergence.

MBGD performs 5,000 updates per epoch, one for each training house, giving fast progress but highly noisy convergence.