FuXi: Becoming a Full-fledged Logical Reasoning System

[by Chimezie Ogbuji]

I've been doing alot of "Google reading" lately

Completing Logical Reasoning System Capabilities

With the completion (or near completion) of PML-generating capabilities for FuXi, it is becoming a fully functional logical reasoning system. In "(Artificial Intelligence: A Modern Approach)" Stuart Russel and Peter Norvig identify the following main categories for automated reasoning systems:

  1. Theorem provers
  2. Production systems
  3. Frame systems and semantic networks
  4. Description Logic systems

OWL and RDF are coverage for 3 and 4. The second category is functionally covered by the RETE-UL algorithm FuXi employs (a highly efficient modification of the original RETE algorithm). The currently developing RIF Basic Logic Dialect covers 2 - 4. Proof Markup Language covers 1. Now, FuXi can generate (and export visualization diagrams) Proof Markup Language (PML) structures. I still need to do more testing, and hope to be able to generate proofs for each of the OWL tests. Until then, below is a diagram of the proof tree generated from the "She's a Witch and I have Proof" test case:

# http://clarkparsia.com/weblog/2007/01/02/burn-the-witch/
# http://www.netfunny.com/rhf/jokes/90q4/burnher.html
@prefix : <http://www.w3.org/2000/10/swap/test/reason/witch#>.
@keywords is, of, a.
#[1]    BURNS(x) /\ WOMAN(x)            =>      WITCH(x)
{ ?x a BURNS. ?x a WOMAN } => { ?x a WITCH }.
#[2]    WOMAN(GIRL)
GIRL a WOMAN.
#[3]    \forall x, ISMADEOFWOOD(x)      =>      BURNS(x)
{ ?x a ISMADEOFWOOD. } => { ?x a BURNS. }.
#[4]    \forall x, FLOATS(x)            =>      ISMADEOFWOOD(x)
{ ?x a FLOATS } => { ?x a ISMADEOFWOOD }.
#[5]    FLOATS(DUCK)
DUCK a FLOATS.
#[6]    \forall x,y FLOATS(x) /\ SAMEWEIGHT(x,y) =>     FLOATS(y)
{ ?x a FLOATS. ?x SAMEWEIGHT ?y } => { ?y a FLOATS }.
# and, by experiment
# [7]   SAMEWEIGHT(DUCK,GIRL)

Shes a witch and I have proof trace

There is another test case of the "Dan's home region is Texas" test case on a python-dlp Wiki: DanHomeProof:

@prefix : <gmpbnode#>.
@keywords is, of, a.
dan home [ in Texas ].
{ ?WHO home ?WHERE.
  ?WHERE in ?REGION } => { ?WHO homeRegion ?REGION }.

Dans home region is Texas proof

I decided to use PML structures since there are a slew of Stanford tools which understand / visualize it and I can generate other formats from this common structure (including the CWM reason.n3 vocabulary). Personally, I prefer the proof visualization to the typically verbose step-based Q.E.D. proof.

Update: I found nice write-up on the CWM-based reason ontology and translations to PML

So, how does a forward-chaining production rule system generate proofs that are really meant for backward chaining algorithms? When the FuXi network is fed initial assertions, it is told what the 'goal' is. The goal is a single RDF statement which is being prooved. When the forward-chaining results in a inferred triple which matches the goal, it terminates the RETE algorithm. So, depending on the order of the rules and the order that the initial facts are fed it will be (for the general cases) less efficient than a backward chaining algorithm. However, I'm hoping the blinding speed of the fully hashed RETE-UL algorithm makes up the difference.

I've been spending quite a bit of time on FuXi mainly because I am interested in empirical evidence which supports a school of thought which claims that Description Logic based inference (Tableaux-based inference) will never scale as well the Logic Programming equivalent - at least for certain expressive fragments of Description Logic (I say expressive because even given the things you cannot express in this subset of OWL-DL there is much more in Horn Normal Form (and Datalog) that you cannot express even in the underlying DL for OWL 1.1). The genesis of this is a paper I read, which lays out the theory, but there was no practice to support the claims at the time (at least that I knew of). If you are interested in the details, the paper is "Description Logic Programs: Combining Logic Programs with Description Logic" and written by many people who are working in the Rule Interchange Format Working Group.

It is not light reading, but is complementary to some of Bijan's recent posts about DL-safe rules and SWRL.

A follow-up is a paper called "A Realistic Architecture for the Semantic Web" which builds on the DLP paper and makes claims that the current OWL (Description Logic-based) Semantic Web inference stack is problematic and should instead be stacked ontop of Logic Programming since Logic Programming algorithm has a much richer and pervasively deployed history (all modern relational databases, prolog, etc..)

The arguments seem sound to me, so I've essentially been building up FuXi to implement that vision (especially since it employes - arguably - the most efficient Production Rule inference algorithm). The final piece was a fully-functional implementation of the Description Horn Logic algorithm. Why is this important? The short of it is that the DLP paper outlines an algorithm which takes a (constrained) set of Description Logic expressions and converts them to 'pure' rules.

Normally, Logic Programming N3 implementations pass the OWL tests by using a generic ruleset which captures a subset of the OWL DL semantics. The most common one is owl-rules.n3. DLP flips the script by generating a rule-set specifically for the original DL, instead of feeding OWL expressions into the same network. This allows the RETE-UL algorithm to create an even more efficient network since it will be tailored to the specific bits of OWL.

For instance, where I used to run through the OWL tests in about 4 seconds, I can now pass them in about 1 secound using. Before I would setup a RETE network which consisted of the generic ruleset once and run the tests through it (resetting it each time). Now, for each test, I create a custom network, evaluate the OWL test case against it. Even with this extra overhead, it is still 4 times faster! The custom network is trivial in most cases.

Ultimately I would like to be able to use FuXi for generic "Semantic Web" agent machinery and perhaps even to try some that programming-by-proof thing that Dijkstra was talking about.

Chimezie Ogbuji

via Copia
1 response
If you're referring to RETE-UL as the most efficient pattern matching algorithm, that claim is only true if all facts fit Robert Doorenboss' IAV design. Mapping EAV to RETE-UL triple results in grossly inefficient RETE network. The most optimal design for entity objects is Dr. Forgy's design, not RETE-UL.



I believe I went to great lengths to show mathematically why EAV-triple design is the wrong approach for entity objects with n-tuples.



peter