LoboVault Home

Three algorithms for causal learning


Please use this identifier to cite or link to this item: http://hdl.handle.net/1928/12107

Three algorithms for causal learning

Show full item record

Title: Three algorithms for causal learning
Author: Rammohan, Roshan Ram
Advisor(s): Luger, George F.
Committee Member(s): Caudell, Thomas Preston
Williams, Lance R
Reda Taha, Mahmoud
Stern, Carl
Department: University of New Mexico. Dept. of Computer Science
Subject: artificial intelligence, machine learning, causal learning, automated experimentation, Bayesian networks
LC Subject(s): Artificial intelligence.
Machine learning.
Bayesian field theory.
Degree Level: Doctoral
Abstract: The field of causal learning has grown in the past decade, establishing itself as a major focus in artificial intelligence research. Traditionally, approaches to causal learning are split into two areas. One area involves the learning of structures from observational data alone and the second, involves the methodologies of conducting and learning from experiments. In this dissertation, I investigate three different aspects of causal learning, all of which are based on the causal Bayesian network framework. Constraint based structure search algorithms that learn partially directed acyclic graphs as causal models from observational data rely on the faithfulness assumption, which is often violated due to inaccurate statistical tests on finite datasets. My first contribution is a modification of the traditional approaches to achieving greater robustness in the light of these faults. Secondly, I present a new algorithm to infer the parent set of a variable when a specific type of experiment called a `hard intervention' is performed. I also present an auxiliary result of this effort, a fast algorithm to estimate the Kullback Leibler divergence of high dimensional distributions from datasets. Thirdly, I introduce a fast heuristic algorithm to optimize the number and sequence of experiments required towards complete causal discovery for different classes of causal graphs and provide suggestions for implementing an interactive version. Finally, I provide numerical simulation results for each algorithm discussed and present some directions for future research.
Graduation Date: December 2010
URI: http://hdl.handle.net/1928/12107

Files in this item

Files Size Format View Description
Rammohan_Dissertation_Signed.pdf 12.79Mb PDF View/Open Complete and Signed Dissertation

This item appears in the following Collection(s)

Show full item record

UNM Libraries

Search LoboVault


My Account