IW Meeting 2010-09-02

From Inference Web

Jump to: navigation, search

Contents

Meeting Info

IW Overall Objectives

     1) PML ontology: (OWL level only) potential non-monotonic  changes (Li) 
     2) PML practice  
         2.1) Use cases and examples (Tim) 
         2.2) RDF level issues (Cynthia) 
         2.3) PML Tools / API issues (Cynthia) 
         2.4) PML primer (Tim) 
     3) PML and OPM mapping (Paulo) 
     4) IW management (TBD)
        4.1) inference web infrastructure and research plan
        4.2) technologies for engagement 
     5) PML Graphical Notation

Agenda

  • (Geoff) combine proof
   Can you bring me up to speed with the status of your work on proof combining ... 
   has there been any progress since we last left off? 
   Is there still interest, or has the group moved on to other things?

summary of summer meetings

(li) management
  • put anything need to be decided in near future in issue tracker.


PML ontology


PML tools
  • How will the current PML tools be affected by naming Information with the value of pmlp:hasURL and omitting pmlp:hasURL?
  • PML API Documentation, how that related PML Primer


Tim
PML uses cases
  • CSV2RDF converter process
  • Incubator Group Scenarios (first one done)
  • six types of existing PML examples (need PML files and a homepage)
   1) proofs that demonstrate our capability of verifying PML. That was done by pryiandra and it has PPDR (or inference ML) stuff in it;
   2) proof involving typicality assumptions in support of NIMD;
   3) JTP proofs derived from IBM UIMA logic;
   4) combined proofs of JTP reasoning and UIMA-extracted knowlege; mississipi is just UIMA-extracted knowledge. We used to have examples of COMBINED proofs. proofs is 4 are very large (probably the largest proofs we had at that time)
   5) proofs involving the use of tactics and declarative rules (possible also with the help of Inference ML)
   6) proofs involving tasks such as proofs derived by SRI SPARK; I am talking about CALO and Alyssa may have access to some of these proofs 


Attendees

  • Cynthia Chang
  • Li Ding
  • Tim Lebo
  • Paulo Pinheiro da Silva
  • Deborah McGuinness
  • Geoff

Notes

Geoff: The QMS part was not a huge success - the room was not right for a productive discussion, and the crowd was interested in other issues raised during the preceding EMS panel. However, the discussion definitely provided some ideas we can use. Here are my notes ...

+ A good proof does not depend on specific reasoners. Solutions should
  use a set of primitive inferences.
+ Metrics
 - Time/space of replaying a proof
 - Use (or lack of) use of definitions
 - Size of the proof object
+ Proofs should use structure sharing
+ Symbols in solutions must be unique and unambiguous
+ Proofs can be compressed using hashcons, or Larry Wos' work
+ Proofs that can be represented at different levels of granularity are
   useful for different applications.
+ For answer set programming there the problem sepcfies a quality metric
  that system optimise, so ranking is simple.
+ In hardware model checking the solutions are very hard to print, and thus
  solutions are very valuable.
+ Proof that are minimal, or use minimal inference pathways, are good (Ian
  Horrocks - I didn't quite get the idea).
+ Evaluation of solutions is tricky, and could be a separate field of study.
+ Solutions are big, and too many of them could swamp us.
  • Paulo Pinheiro da Silva: i am wondering what are the requirements for pml 2.1 regarding proof combination
  • Paulo Pinheiro da Silva: it appears that most of the proof combination work is on how to post-process proofs rather than new ways of capturing and encoding provenance
  • Geoff: Right
  • Paulo Pinheiro da Silva: it appears that PML is a good fit for proofs generated by answer set programming (I have discussed this with people at NMSU) but we do not have a single example to demonstrate this.
  • changcs1: i always think pmlj:InferenceStep.fromAnswerOrQuery maxCard=1 can be relaxed. in combined proofs, one infStep can be used for diff answers.

[*Geoff: We will proceed with existing PML, doing hill-climbing with single combinations, using generalized clustering coefficient as the heuristic. Geoff will send Cynthia the GCC program. Cynthia will work on the current implementation to do this. Geoff and Cynthia to liase on progress and report next week.

Personal tools
Navigation