Suppose you trained SVM classifier with radial kernel and suspect that it underfits the data (i.e. you have high bias and low variance: the test error is very close to the training error, but both are quite large). If you would like to obtain a better fit to the training data (increase the model's variance, decrease the bias), you should try: *Note, that there is some confusion in the notation, in case you are reading additional sources: we defined A as the limit (allowance) on the sum of misclassification errors e; (and your book defined it as C). The 'cost' parameter C in svm() function is the inverse of that: large A means we are OK with many mis-classifications i.e. their "cost" (penalty associated with them) is low and vice versa, low allowance A means very high cost of each misclassification. In the literature, the notation C is predominantly used for such *cost" (same way we did it in the slides), so be careful to understand first what your source is talking about! The possible answers to this question are formulated in terms of allowance A as we defined it (what your textbook called "C"): large A = sum of e; can be large (and correspondingly A is also large). Increasing A = decreasing the misclassification 'cost'. O increasing y, increasing A O decreasing y, increasing A O increasing y, decreasing A O decreasing y, decreasing A