Handbook of Perception and Cognition, Vol.14 Chapter 4: Machine Learning

Machine learning is the subfield of AI concerned with intelligent systems that learn. To understand machine learning, it is helpful to have a clear notion of intelligent systems. This chapter adopts a view of intelligent systems as agents — systems that perceive and act in an environment; an agent is intelligent to the degree that its actions are successful. Intelligent agents can be natural or artificial; here we shall be concerned primarily with artificial agents. Machine learning research is relevant to the goals of both artificial intelligence and cognitive psychology. At present, humans are much better learners, for the most part, than either machine learning programs or psychological models. Except in certain artificial circumstances, the overwhelming deficiency of current psychological models of learning is their complete incompetence as learners. Since the goal of machine learning is to make better learning mechanisms, and to understand them, results from machine learning will be useful to psychologists at least until machine learning systems approach or surpass humans in their general learning capabilities. All of the issues that come up in machine learning — generalization ability, handling noisy input, using prior knowledge, handling complex environments, forming new concepts, active exploration and so on — are also issues in the psychology of learning and development. Theoretical results on the computational (in)tractability of certain learning tasks apply equally to machines and to humans. Finally, some AI system designs, such as Newell's SOAR architecture, are also intended as cognitive models. We will see, however, that it is often difficult to interpret human learning performance in terms of specific mechanisms. Learning is often viewed as the most fundamental aspect of intelligence, as it enables the agent to become independent of its creator. It is an essential component of an agent design whenever the designer has incomplete knowledge of the task environment. It therefore provides autonomy in that the agent is not dependent on the designer's knowledge for its success, and can free itself from the assumptions built into its initial configuration. Learning may also be the only route by which we can construct very complex intelligent systems. In many application domains, the state-of-the-art systems are constructed by a learning process rather than by traditional programming or knowledge engineering. 1 Machine learning is a large and active field of research. This chapter provides only a brief sketch of the basic principles, techniques and results, and only brief pointers to …