A target function expresses the desired input-output behavior of an intelligent task. It is usually written as y = f(x): x represents input variables, and y represents the output produced for a particular input. In many real-world problems, the underlying function is unknown and may be arbitrarily complicated, so a model must approximate it rather than receive it as an explicit design. section 7.1 section 7.2
Inputs can be numeric scalars, vectors, matrices, or other collections of variables. Variables may be continuous, such as a stock price, or categorical, such as binary expert advice. The output is likewise numeric and represents the solution for that input instance. section 7.2
Stock-selling decision
A stock-selling target function takes inputs such as purchase price, current price, and expert advice, then returns 0 for “do not sell” or 1 for “sell.” A model that approximates this function provides a mechanical decision for each set of input values. section 7.2
Neural networks provide a common framework for approximating many different target functions. Individual neurons represent relatively simple functions; a network combines them through weighted connections. Its expressive power depends on the number of neurons, connections, and weights, and more complicated target functions require greater expressive power. section 7.1
