The drawback of the perceptron learning rule introduced in the previous chapter is that it requires the learning pattern to be linearly separable, Linearly separable i.e., there must be a linear discriminant function that makes misclassification zero. For learning patterns that are linearly nonseparable, Linearly nonseparable the weight modification procedure is repeated infinitely, and no solution can be reached. Even if we forcefully terminate the process because there is no possibility of convergence, there is no guarantee that the weights obtained at that time will be optimal. In general, it is difficult to confirm in advance whether or not linear separation is possible. This chapter introduces a general learning algorithm that can be applied to the case where linear separation is not possible.