Cognitive Networks: Towards Self-Aware Networks

5.2. Problem Formulations in Machine Learning

Treatments of machine learning (e.g., ^{[] []}) typically organize the field along representational lines, depending on whether one encodes learned knowledge using decision trees, neural networks, case libraries, probabilistic summaries, or some other notation. However, a more basic issue concerns how one formulates the learning task in terms of the inputs that drive learning and the manner in which the learned knowledge is utilized. This section examines three broad formulations of machine learning.

5.2.1. Learning for Classification and Regression

The most common formulation focuses on learning knowledge for the performance task of classification or regression. Classification involves assigning a test case to one of a finite set of classes, whereas regression predicts the case's value on some continuous variable or attribute. In the context of network diagnosis, one classification problem is deciding whether a connection failure is due to the target site being down, the target site being overloaded, or the ISP service being down. An analogous regression problem might involve predicting the time it will take for the connection to return. Cases are typically described as a set of values for discrete or continuous attributes or variables. For example, a description of the network's state might include attributes for packet loss, transfer time and connectivity. Some work on classification and regression instead operates over relational ...