On the Bayesian Approach to Learning
Silvia Acid, Luis M. de Campos, Antonio González, Rafael Molina, Nicolás Pérez de la Blanca · 1992
The aim of this work is to introduce CASTLE (Causal Structures from Inductive Learning), a tool based on the bayesian approach to learning. CASTLE can be used so far to learn causal structures from raw data, propagate knowledge throughout polytrees, simulate and also edit polytree dependent distributions. CASTLE ([1] and [2]), is currently being developed by the authors in the DECSAI at the University of Granada. Basically, CASTLE estimates, from a file of examples, the (in)dependencies among the variables involved in the examples in order to build a polytree displaying such (in)dependencies. The steps to construct such polytree are: setting constrains among the variables, selecting a criterion to calculate the skeleton of the polytree and, finally, selecting the criterion to direct the obtained skeleton. Once the polytree is built, CASTLE allows the user to propagate knowledge throughout the obtained singly-connected graph using what has been called the bayesian approach to the knowledge propagation task. CASTLE can be also used as a platform where to test learning methods since it allows the users to create polytrees and simulate data from them and use the generated sample as learning samples.