In this section we have described numerous estimates that can be obtained by using locally weighted learning for function approximation. One question to ask is what are all these estimates good for? Although learning algorithms are almost always described in terms of their ability to make predictions, that is not their purpose. A good learning algorithm should help us make decisions. If a friend calls you up and says ``I think this stock will double in price this year.'', do you immediately go out to purchase the stock? Its more likely that you reply with ``How sure are you that this will happen?'' Making decisions based on the advice of learning algorithms is the same way.
We have already observed that optimization algorithms can make use of these estimates. Confidence intervals are also useful for computing the expected gains or losses of potential decisions. Some possible outputs may represent catastrophic consequences and any decision that has a significant chance of yielding those results should be avoided. Later on in this tutorial we will see more examples of confidence intervals being used in decision making.