A half-century survey on The Universal Turing Machine
Models of incremental concept formation
Artificial Intelligence
A learning as the evolution of representation
A learning as the evolution of representation
Empirical substructure discovery
Proceedings of the sixth international workshop on Machine learning
A theory of natural learning
A Theory of Program Size Formally Identical to Information Theory
Journal of the ACM (JACM)
Explanation-Based Generalization: A Unifying View
Machine Learning
Explanation-Based Learning: An Alternative View
Machine Learning
Failsafe: a floor planner that uses EBG to learn from its failures
IJCAI'87 Proceedings of the 10th international joint conference on Artificial intelligence - Volume 1
Algorithmic information theory
IBM Journal of Research and Development
Hi-index | 0.00 |
We present an approach to unsupervised concept fonnation, based on accumulation of partial regularities. Using an algorithmic complexity framework, we defme regularity as a model that achieves a compressed coding of data. We discuss induction of models. We present induction of finite automata models for regularities of strings, and induction of models based on vector translations for sets of points. The concepts we develop are particularly appropriate for natural spaces - structures that accept a decomposition into recurrent, recognizable parts. They are usually hierarchical and suggest that a vocabulary of basic constituents can be learned before focussing on how they are assembled. We define and illustrate: Basic regularities as algorithmically independent building blocks of structures. They are identifiable as local maxima of compression as a function of model complexity. Stepwise induction consists in finding a model, using it to compress the data, then applying the same procedure on the code. It is a way to induce, in polynomial time, structures whose basic components have bound complexity. Libraries are sets of partial regularities, a theoretical basis for clustering and concept formation. Finally, we use the above concepts to present a new perspective on explanation based generalization. We prove it to be a language independent method to specialize the background knowledge.