Automated Causal Analysis

This project considers how to partially automate the testing of ergalic theories. We are considering how to represent ergalic theories as a family of causal models, how to determine the efficacy of each experiment or collection of experiments, and how to update the current model with the results of experiments. The goal is to work towards "closing the loop" on the entire process of model testing. This work draws upon results and research in machine learning, artificial intelligence, and philosophy of science.

The mathematical language of structural causal models (SCMs) [Pearl 2009] can be used as a model framework. The theory underlying SCMs is the product of several traditions, including the development of structural equation models [Bollen 1989], and the parallel development of the use of graphs (or graphical models) to represent informational relationships and relevance, in particular as a representation of statistical dependence and independence [Koller 2009, Lauritzen 1996, Pearl 2009].

SCMs provide a natural representation for common empirical modeling tasks. SCMs make causal claims explicit: the presence of a directed link expresses that there is a causal influence, an influence that may be tested if the appropriate observational data is available or we can intervene on particular variables.

In this project we are developing AMELIE, a cyber workspace that provides computational support for causal modeling throughout the process of an empirical investigation.

AMELIE connects to a variety of experiment domains that can perform experiment workflows produced by AMELIE. Each such experiment run produces result data that is fed back into AMELIE to analyze to determine the extent to which that data supports the hypotheses tested by that experiment, possibly adding support for the underlying structural causal model.

• ANTARES Experiment Domain (with John Kececioglu, Thomas Matheson, and Abhijit Saha)
• AZDBLab Experiment Domain (with Sabah Currim, with the experiment manager being the "Arizona Database Laboratory", or the AZDBLab system)
• Biological Signal Transduction Network Experiment Domain (with Ryan Gutenkunst)
• UA Academic Experiment Domain (with Sabah Currim and Guillermo Uribe, with the experiment manager being "Student Prediction of Courses," or the SPOCK System)
• Validation Experiment Domain (with David Sidi, with the experiment manager being the "Heuristic Artificial World Kreator" System, or HAWK, with the scientist's interface being SHAWK) )

Young Suh, Richard T. Snodgrass, and Rui Zhang, "AZDBLab: A Laboratory Information System for a Large-Scale Empirical Study," VLDB Demos, September, 2014.

Thomas Matheson, Abhijit Saha, Richard T. Snodgrass, and John Kececioglu, "ANTARES: A Prototype Transient Broker System," in American Astronomical Meeting Abstracts, 223, #343.02, 2014. (abstract)

Thomas Matheson, Abhijit Saha, Richard T. Snodgrass, and John Kececioglu, "ANTARES: The Arizona-NOAO Temporal Analysis and Response to Events System," in Hot-Wiring the Transient Universe 3}, P. Wozniak, M. Graham, and A. Mahabal (eds.), 2014. (pdf)

Clayton Morrison and Richard T. Snodgrass, "Computer Science Can Use More Science," Communications of the ACM, June 2011, pp. 36-39. (PDF).

A graphic story that appeared January 8, 2014 on the NOAO web site shows the development and promise of the ANTARES system. (Click on the thumbnail below to see it in all of its glory.)

"How to Find the Rarest of the Rare in Southern Skies." article in UANews, December 17, 2013

"Anything you can imagine happening in our vast universe already has," article in the Arizona Daily Star, October 18, 2013