You are on page 1of 3

,r<-

Performance measures of .rtificiaf Inielligence ded-uction - pl-ans based" cn Chaitin I s notion of coniplexit;y.
Rago Francesco

Cassa d.i Ri-sparmio i Puglia

Eari

lhe main purpose of this work is te apply the d.efinitien c information measure of Kolgomoroff (f) aruC. Chaiiin (2) to ihe praductions of rrrtificial Intel]igence ded.uetion plans in terms of their und-erlyjng structure - In ths way it is psssible to stu* d-y their internal structure using tne concept cf mutua.J- infcrma* tion noting'bhat ther.e is a close relationship bettveen the co:* c*pts of d*duction plans complexity and information j-n tl:e
Shannon serlse
*

of certain d"i"ree ted- graphs salled Ced-uciion plans. T'he s*ructu.re of a d.educticn plan fcr a set C of ciaussr '*.thich have a set *f, truth valueo * Ss a rcoted. tree * Every nod.e , exeept the root , is s. cCIpy of a variant of claus es cf S -. For more d.etails s ee ( 3 ) and" (+)..Farmally let S a set *f clauses* -4. pair (g,f) is a ded-uobion plan for S where G is a graph anci. f is a mep of..E(G), called the labelin$ " If we iJ.se a set d"ef intion of logic r ws oen &ssociate tc the 6raph of the Ced.uction plan the gensrative graph of a Sr&"m$&.r. Tt s possihle tc build a set of prod"uctions of the
" d-eeluntion

sysiem rel-j-es or3 the oonstruction

fype;

i=I1."c;11

j=Irea,rki

where

of proouctions that A, bas as a prenise. o_-eactr-prod-uctioo it is .assigned a coeworri 'of .the.d"ed.uction.- rules defined by the language derived. by the ded,ution plan. - + To each production of G it is al-so assigned a codeword. zi5e{Orf1 .
number

k, is the

f'or
o\ O-i

a.ll i=Itr."*3rr
rt lJ.

and.

'

b) Ilz.. = I ".' l-J I t al.l e) los ki(= | *t;


!

l_- -(J

n {J.-

l^I

ki= ; we 6etl iff i=i an j for jrl = Io.


where

t-L

; -

f?"cGcl!.

l*

'1.

. c t Li\; l-

Ia

to* ki + r

Tr-Lt.colI'^-;
-..!_

If we have a leftnost d.erivaticn (4 or4ir..,!'u{t) of'd i e:iii4i, Vie {fr...rt] ,belonging io the finite sets of nonierminal-s and. terninals, and4 i_IFr(, tlren we defi-ne the associated- cod-eword-

p(dr) as

z(c(r_1*41). Each d.erivatlon of G correspond.s to a particular path through the graph, starting ai its root and ending at the nod.e presenting the last elenent of the derivation. If we 1abe1 each branch by the cod-eword. assi-gned- to the productionreach path in the grapb is conpletely d.escribed. by the cod.eword assoeiated. with leftrnost d-erivation belonging to that path. Let be y a sentence and y(i) the i-th leftmost d.erir.atic::
t(4*)=z(a ol-d.r)z(c(rLC)z ..

of y with codeword- ni(f ) . Using Chai-tint s cond.i-tional corpl ::r:ty, of a strinB,the complexity is C(y)=mini C(y(i)) where c(y(i.))=l ni(r)l - rf we have two sentences try the conditio"ai ccnplexity i-s C(t,/y)=min. rjC(t(j)/v(i)). The information H(y) of a sentence y is defined. as

il:e conditional infornqation cf two sentences y and. t is: H(t/v)=mini5 ( c(t( j ) /v(i) )-c(y(i)/v(i) ) ). The mutual i-nrormation of two sentences y and" t is : if(y)=C(y)
ano"

)) j-nformation gives values tr'or d.ifferent sentences the nutual that are measures of the s j-milarity between the graph of the d.eCr:ction plans and- so it is possible to have a performance ind.ex related. to internal structure. ft is possible to infer the possibility of monitoring an artificial intelligence machj-ne.
Ref erenc es

r(t its)= *"*j-,j (e(t(j ))-c(t(j ) /v(i))+c(y(i)/y(i)

to the d.efinition of the concept of the anount of inf ornationr rtProbl. inforn. transn. voI.1 1965 (2) C.,f . Chaitinr'rA theory of prograns size id-entical to informr ation theory, "J. Ass. Conut. I{ach. n. t A975 (, ) P.T. Cox, rrDeduction P1ans, rPh.d.- d.issertation, Univ. lJaterloo,
Konogorof, " Three approaches
1977

(I) A.N-

(4) J.A. RobinsonrrtA nachiac-oriented logic based on the resolution principle", J. Ass. Corap. i'iach.rvol. 12 Jan. T965

You might also like