Version Space

This is an excerpt from the content

Definition

Mitchell (1977, 1982) defines the version space for a learning algorithm as the subset of hypotheses consistent with the training examples. That is, the hypothesis language is capable of describing a large, possibly infinite, number of concepts. When searching for the target concept, we are only interested in the subset of sentences in the hypothesis language that are consistent with the training examples, where consistent means that the examples are correctly classified (assuming deterministic concepts and no noise in the data). While the version space may be infinite, it can often be represented in a compact manner by maintaining only its bounds, the most specific ( Most Specific Hypothesis) and most general hypotheses. Any hypothesis that is more general than a hypothesis in the most specific bound and more specific than a hypothesis in the most general bound is in the version space.

Cross References

Learning as Search

Noise