Prior Analytics
9 July 2010, 19:23 UTCIncoherent Prediction Markets?
I have known about prediction markets for some time, and find botboth the theoretical arguments and the evidence convincing. But only today did I register for one and look around. I found something odd.
read more...
22 June 2010, 2:15 UTCWho Knows Best?
Ann Nicholson just pointed me to an excellent discussion of two approaches to behavioral economics: nudge vs. regulate.
read more...
22 June 2010, 1:13 UTCEmacs Calendar & pdfLaTeX
How to print emacs calendar landscape using pdfLaTeX.
read more...
13 April 2010, 1:12 UTCRiding the E+ Sun Ray EZ Recumbent
6 April 2010, 18:46 UTCPerformance-Weighted Opinions
27 March 2010, 1:00 UTCYudkowsky on Priors
2 February 2010, 19:28 UTCFundamental Law
14 January 2010, 2:39 UTCPhoto in The Fairest
19 November 2009, 21:47 UTCProbability Words
6 November 2009, 15:12 UTCExplosives Divining Rod
6 November 2009, 10:58 UTCConsistency
31 October 2009, 0:57 UTCPHIL 721 Advanced Seminar: Causation
22 October 2009, 3:29 UTCBNs for Intelligence Analysis
21 October 2009, 23:59 UTCBNs for Initiative Assessment
6 October 2009, 13:33 UTCNetbooks vs Thinkpad X40
14 September 2009, 19:12 UTCpasswd:
13 August 2009, 0:37 UTCPay Too Much For Food & Power
6 November 2008, 16:51 UTCGelman on the Election
9 October 2008, 3:32 UTCRail vs. Bus
5 August 2008, 12:12 UTCBicycle Helmets & Pedestrian Casualties
9 July 2008, 14:58 UTCReston to GMU Bike Route -- Updated
1 July 2008, 21:45 UTCCorrelation & Causation
18 June 2008, 20:38 UTCThe Crowd Within
16 June 2008, 19:40 UTCAudiophiles, Significance Tests, Greenspun's Tenth Rule
10 June 2008, 13:27 UTCDemocracy
13 March 2008, 0:24 UTCKrugman on Interstellar Trade
6 March 2008, 1:33 UTCScience & Magic
29 November 2007, 1:58 UTCFixed Links
10 October 2007, 16:41 UTCAviation collision experiments
22 September 2007, 14:52 UTCPay for free software
We presented
"Rapid Initiative Assessment for Counter-IED Investment" at the 7th Annual
Bayesian Applications Workshop and the
77th MORS Symposium, both on June 18, 2009. One of the more interesting aspects for me was working on
link strength graphs.
I read
Imme Ebert-Uphoff's link-strength work, which referred to the paper that Ann Nicholson and Nathalie Jitnah were finishing just as I arrive at Monash. After considering the two, I decided to use another measure that Nicholson and Jitnah had used, which I called Cut Link Strength.
Another approach is to compare with and without
the link . This was the “gold standard” that
Nicholson & Jitnah (1998) used to evaluate their (linkstrength-like) approximate inference. But we can afford to
use the gold standard itself.
When cutting the arc , we marginalize over X, which leaves unchanged the marginal distribution for Y. However, if the arc was not completely superfluous, the new P(Y|x) will differ from the old for at least some .
To control for possible back paths like , we use an intervention operator “||” rather than a regular
conditioning operator “|”. (An intervention operator, often
called “do(x)”, blocks backwards inference, effectively
cutting the links into X.) Let be the resulting
distribution in the original graph, and let be the
same in the new graph, with . The link strength is
the expected distance between these two distributions.
Thanks to Ebert-Uphoff for his Matlab code. I translated it for Quiddity and then added the Cut Link Strength method. I also worked on a Netica version.
When testing the method, I noticed that when you delete an arc in Netica
, Netica marginalizes Y as if X were in its
first state. This only makes sense if that is somehow a preferred default (as it might be in binary nodes). Otherwise it is quite arbitrary.
The talks were nearly simultaneous, so we had to split up. Kathy, Kellen, & I went to Fort Leavenworth while Ed went to Montreal. To put a positive spin on it, we got the better conference center, by far. And Kathy & I did get to join him in Montreal the next day for the UAI conference. Alas, the cab ride out showed that I had missed the most interesting parts of the city. But PSA'2010 is in Montreal -- time to work on my paper!
Thanks to our sponsors, both for funding the work and clearing the paper for publication.