IW Meeting 2011-03-17

From Inference Web

Jump to: navigation, search

Contents

Meeting info

Agenda

  • Catching up!

Li is also at this meeting so he will also miss this meeting. we will still be in australia next thurs so both of us will not be on for 2 weeks.

on the bright side though - my keynote is on provenance and next week's visit should make progress on our provenance work with csiro.

Geoff proof combo

greedy algorithm by hand. results for two proofs (one in set domain). results correct but disappointing. nothing fantastic.

Geoff looked at results and the results look correct.

next: 1) automating (Cynthia) (TODO: Geoff to answer Cynthia's question for how to automate) 2) TODO: geoff find more candidate problems.

process control

PXTP paper: need one example that demonstrates success - two hand examples don't cut it.

by hand: 140+1 140+2

Cynthia: only thing by hand: of all numbers, she chooses one. Then the selected one goes into combining program. loop of select one, run combine program -- select another, run combine program, repeat.

NOTES from last week (Cynthia) [11:21:10 AM] changcs1: hill climbing is being used in combined proofs. the A-B proofs result is out. next we are working on real proofs. SEU140+1 will be done today. then SEU140+2... one question I have is when you have multiple combined proofs having the same gcc numbers, which one we should choose for the next run (climbing further). [11:30:25 AM] deborah_mcguinness: we might get a status update from geoff on the may paper - both where he is and what he needs from us. [11:42:21 AM] liding2006: if two proofs have the same number, they should meet the ultimate goal [11:44:55 AM] liding2006: is gcc number is correct? [11:45:09 AM | Edited 11:45:12 AM] liding2006: is gcc number good enough? [11:48:18 AM] liding2006: the key question is "how many proofs have the best gcc number in average?" [11:50:44 AM | Edited 11:50:54 AM] liding2006: which proofs have been returned to geoff [11:54:41 AM] liding2006: in general, replace sub-tree [11:55:31 AM] liding2006: two difference [11:56:26 AM] liding2006: 1. different measure, size of tree; gcc of tree [12:04:49 PM | Edited 12:04:59 PM] liding2006: todo: clarify the problem and solution -- li [12:06:16 PM | Edited 12:06:49 PM] liding2006: todo: more measureable problems [12:10:01 PM] liding2006: generalizable hypotheses [12:10:42 PM] liding2006: e.g. categorize cobmined proofs [12:12:39 PM] liding2006: e.g. generate random sample combined proofs, how to gurantee it [12:13:59 PM | Edited 12:14:51 PM] liding2006: e.g summarize a collection of alternative proof, proof clustering [12:17:29 PM | Edited 12:18:17 PM] liding2006: e.g. declare our linked proofs in LOD cloud, claim number of triples, generate a statistical page [12:33:37 PM] liding2006: http://data-gov.tw.rpi.edu/ws/lodcx.php


Catching up

Paulo's meeting

latest: connecting sci data to publications (using PML) new issues.

e.g. publications

a lot of data about publications (bibtex, bibliographies) - lots of data.

advantage of PML for publication

paper with many authors - what is contribution of each author. similar to our previous wikipedia PML work (revisions)

did NOT do: combining one article w.r.t. another article.

e.g. two articles of wikipedia. how does one paper evolve with considering other articles.

Pa with Ax and Pb with Ay

Pa :cites Pb

part of Pa proof has Pb antecedent. <--- the challenge

Since proofs are combined, Ay would appear to be author of Pa (but that 's not TRUE)


also: paper basing assertions on data - and data that feeds into that data.


also: co-citing paper.

problem with PML justification loop: loading it. when reasoning with it.


Tim: example? Paulo: Geoff's combining proof stuff will need to consider loops. What does it mean for something to be derived from something that derives from itself?


Future IW

3 things:

  • improving the language (layering architecture)
  • improving proof combination (Geoff and Cynthia)
  • appying PML to actually do things

How we use PML:

At UTEP: 1) For automated processes, we use the Data Annotator (DA) 2) For manual processes, we use DerivA 3) Graduate students

both DA and DerivA are based on the PML API

but they are also based on PML-Network (formally called PML lightweight): this is an additional package that serializes PML that needs to be exchanged in a client-server architecture

At RPI:

4) Tim is using plain Java and shell (but no PML API) 5) Students in class generate some PML (How?) 6) Grad students use PML as well (How?) - Tim's tools for data aggregation 7) Stephan and Patrick use PML (How?)

Some questions: who else is using the PML API? do we see value on the use of PML API?

At Miami:

8) Geoff does not generate PML -- he uses PML that has been converted from TPTP through Cynthia's tools 9) Any student from Geoff is using PML?

What we need:

A) Curated PML FAQ (recipes) - "go-to place" for modeling/generating PML B) People to answer the questions in the FAQ and to clearly connect pointers to tools (PML API, DerivA, DA, Tim's shell code, etc) C) Examples


(a place to put the questions before we have answers) (and we need to classify those questions) (a group of people that look at the questions and try to answer them or approve their solution) Collect some past discussions


http://inference-web.org/wiki/PML_2.1_-_a_Layered_Approach

TODO: Paulo to discuss recipe approach with group. TODO: Tim to set up recipe structure (consider: source code examples. use Ping's examples.)

Personal tools
Navigation