The error metrics are configurable. The two most commonly used are average absolute error (AAE) and mean absolute percentage error (MAPE). You can also specify a moving window so instead of calculating the error at a given point over all data so far, you calculate the metric over the last 1000 records (or whatever you specify).
Swarming doesn't change the model it uses on a per-record basis. Instead, when it picks a new parameter set it runs it all the way through and then takes the final error metric (which may be computed over just the last X records) and compares it to other models tried. On Thu, Jan 15, 2015 at 7:36 PM, Dinesh Deshmukh <[email protected]> wrote: > The documents about swarming illustrate how the error is calculated,but i > cant understand what the error itself mean.That is on what comparisons does > i get the error value? > > What i understand is,if i have 100 data units then swarming would use may > be some 50 data units and then predict 51 using different models and it > choose the model which is close to the prediction of 51(the error > calculation model prediction subtracted by 51). > Finally the best model swarming has given is used to predict unknown data > i.e 101 or 102 or so on... > > Is this view correct? > >
