 |
|
|
|
|
|
|
|
|
|
|
|
|
 |
|
|||||||||||
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
|
 |
 |
|
||||||||||
 |
|
|||||||||||
 |
|
|||||||||||
 |
 |
|
||||||||||
| Seminar Schedule (Seminar Organizer: Prof. Ziv Bar-Joseph) | |
Machine Learning Lunchtime Chats
|
Fall 2005 Seminar
Date: September 16, 2005 Time: 10:00 AM - 11:00 AM Location: 1507 Newell-Simon Hall Speaker: Frederick Eberhardt Masters Candidate Title: Experimental Complexity in Causal Discovery Abstract: In science we are often interested in discovering the causal relations among variables. Interventions are key to this causal discovery since they allow the researcher to distinguish among Markov equivalent graphs. We prove that log(N)+1 experiments are necessary and in the worst case sufficient to discover all the causal relations among N variables, if each experiments consists of a simultaneous and independent randomization on a subset of variables. We assume that there are no latent variables and no feedback, that interventions are possible on all variables and that an experiment returns the conditional independence relations true in the population. The proof specifies a procedure of placing interventions that obtains this bound.