We would like to use SVM to train/predict the covtype data (two-class case). Since earlier results indicate that scaling does not help, we decide to work on the original form.
Finding good parameters is an issue. One set you would like to try first is -c 10000 -g 0.0001. To save the training time, use a large -m and also a larger -e (e.g. 0.01).
To understand the effects of these parameters you may want to play with smaller training/testing subsets. For example, you could use a smaller training set to conduct parameter selection.
You than want to compare your result with those by other methods in earlier homework.
Training this large set is time consuming, so please start your homework as soon as possible.