Chapter 1: Introduction

Some Machine Learning Successes

• Recognizing speech
• Driving an autonomous vehicle
• Classifying new astronomical structures
• Playing world class backgammon

Influences

• Artificial Intelligence
• Bayesian Methods
• Computational Complexity Theory
• Control Theory
• Information Theory
• Neurobiology
• Philosophy
• Psychology
• Statistics

Definition

A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance measure at tasks in T, as measured by P, improves with experience E.

Issues

• What should the training experience be? Should the training experience use direct or indirect examples? Indirect examples mean that the credit assignment problem must be confronted.
• What control does the learner have over the training experience?
• Is the training experience representative of the performance experience?
• What is to be learned? Ideally, we would like to find an operational definition of a target function V that takes a state of the world as input and tells us what to do as output. In practice, V is an approximation.
• How should the target function, V, be represented? There are many choices including rules, tables, weighted linear sums, etc.
• What learning algorithm should be used to modify the target function, V?

Perspective

Machine learning involves searching a very large space of possible hypotheses to determine one that best fits the observed data and any prior knowledge held by the learner.