Learning Analytically and Inductively

Tom Mitchell and Sebastian Thrun

Notice: This paper does not contain an abstract. Here is the Introduction:

Learning is a fundamental component of intelligence, and a key consideration in designing cognitive architectures such as Soar. This chapter considers the question of what constitutes an appropriate general-purpose learning mechanism. We are interested in mechanisms that might explain and reproduce the rich variety of learning capabilities of humans, ranging from learning perceptual-motor skills such as how to ride a bicycle, to learning highly cognitive tasks such as how to play chess.

Research on learning in fields such as cognitive science, artificial intelligence, neurobiology, and statistics has led to the identification of two distinct classes of learning methods: inductive and analytic. Inductive methods, such as neural network Backpropagation, learn general laws by finding statistical correlations and regularities among a large set of training examples. In contrast, analytical methods, such as Explanation-Based Learning, acquire general laws from many fewer training examples. They rely instead on prior knowledge to analyze individual training examples in detail, then use this analysis to distinguish relevant example features from the irrelevant.

The question considered in this chapter is how to best combine inductive and analytical learning in an architecture that seeks to cover the range of learning exhibited by intelligent systems such as humans. We present a specific learning mechanism, Explanation Based Neural Network learning (EBNN), that blends these two types of learning, and present experimental results demonstrating its ability to learn control strategies for a mobile robot using vision, sonar, and laser range sensors. We then consider the analytical learning mechanism in Soar, called chunking, and recent attempts to complement chunking by including inductive mechanisms in Soar. Finally, we suggest a way in which EBNN could be introduced as a replacement for chunking in Soar, thereby incorporating inductive and analytical learning as architectural capabilities.

The following section provides an overview of inductive and analytic principles for learning, and argues that both are necessary for a general learning mechanism that scales up to handle a broad range of tasks. The subsequent section presents the EBNN learning mechanism, together with experimental results illustrating its capabilities. Finally, we consider the general learning mechanism for Soar, and the question of how to best incorporate both inductive and analytic learning within this architecture.

Click here to obtain the full paper (367969 bytes).

@INCOLLECTION{Mitchell95a,
  AUTHOR         = {T. Mitchell and S. Thrun},
  YEAR           = {1996},
  TITLE          = {Learning Analytically and Inductively},
  BOOKTITLE      = {Mind Matters: A Tribute to Allen Newell},
  EDITOR         = {D. Steier and T. Mitchell},
  PUBLISHER      = {Lawrence Erlbaum Associates Publishers}
}