VIEWS: 2 PAGES: 25 POSTED ON: 4/8/2010
Computer Science CPSC 433 - Artificial Intelligence Search Paradigms: Set Based Search Adapted from slides by Jörg Rob Kremer Denzinger ICT 748 kremer@cpsc.ucalgary.ca http://pages.cpsc.ucalgary.ca/~kremer/ 08/04/2010 CPSC 433 Artificial Intelligence 1 Search Paradigms TOC • Introduction • Formal Definitions • Less Formally • Conceptual Example • Designing a set-based search model • Example: Resolution • Example: Blocks world arch • Remarks 08/04/2010 CPSC 433 Artificial Intelligence 2 Search Paradigms Set-based Search Basic Idea: We have a collection of pieces of information (facts) that is (mostly) growing during the performance of a search; a relation between the different pieces is either not known, not of interest or describing only consequences of facts. Represent collection as a set, go from one set to successor by adding/deleting facts according to rules taking into account other facts already in the collection. 08/04/2010 CPSC 433 Artificial Intelligence 3 Search Paradigms Formal Definitions Set-based Search Model Aset = (Sset,Tset): • F: set of facts • Ext {AB | A,B F}: set of extension rules • Sset 2F Power set of F • Tset = {(s,s')| AB Ext A s s'=(s-A)B} Remove some facts and add some facts 08/04/2010 CPSC 433 Artificial Intelligence 4 Search Paradigms Formal Definitions Set-based Search Process Pset = (Aset,Env,Kset): Kset(s,e) = (s-A)B Where • AB Ext Evaluate the “goodness” of a • A s change • A'B' Ext | A’ s fWert(A,B,e) fWert(A',B',e) • AB = fselect({A'B' | A"B" Ext | A" s Break a tie fWert(A',B',e) fWert(A",B",e)},e) For two functions fWert: 2F2FEnvNat and fselect: 22F2FEnv 2F2F 08/04/2010 CPSC 433 Artificial Intelligence 5 Search Paradigms Formal Definitions Set-based Search Instance Insset = (s0,Gset): • s0, sgoal 2F success • Gset(s) sgoal s there is no extension rule applicable in s Failure (there’s nowhere to go) 08/04/2010 CPSC 433 Artificial Intelligence 6 Search Paradigms Less formally • F can consist of solution pieces, solution candidates, parts of a world description, etc. • With Ext we try to get more solution pieces, better candidates, more explicit parts of the description • Or we eliminate wrong pieces, less good solutions, unnecessary explicit parts • We construct the new parts using parts we already have We make implicit knowledge explicit 08/04/2010 CPSC 433 Artificial Intelligence 7 Search Paradigms Less formally • The control selects the extension to apply by – Evaluating each applicable extension into a number (done by fWert) – Considering only extensions with minimal evaluation – Use fselect as tiebreaker • Obviously, there usually are many different fWert and fselect functions • Sometimes fWert can also produce integers or real numbers 08/04/2010 CPSC 433 Artificial Intelligence 8 Search Paradigms Less formally • We start with the given solution pieces, some random solutions, or the given parts of the description (or …) • We stop, if – a complete solution sgoal is part of the actual state or – a good enough candidate that is really a solution is found or – the description is good enough or – a time limit is reached i.e. if enough knowledge (sgoal) is made explicit 08/04/2010 CPSC 433 Artificial Intelligence 9 Search Paradigms Conceptual Example (I):Set-based Search + = + = 3 9 11 + = 08/04/2010 CPSC 433 Artificial Intelligence 10 Search Paradigms Conceptual Example (I): Set-based Search Next state: 08/04/2010 CPSC 433 Artificial Intelligence 11 Search Paradigms Designing set-based search models 1. Identify set of facts F 2. Identify how you create new facts out of known facts (make sure that what you create are really facts!) Ext 3. You have your sets F and Ext that, with our definition earlier, are sufficient to define a set- based search model 08/04/2010 CPSC 433 Artificial Intelligence 12 Search Paradigms Designing set-based search processes 1. Identify possible functions that measure a fact 2. Decide if it is not too computationally expensive to compute the right side of applicable rules a) If it is not too expensive, define fWert by measuring A and B using 1. b) If it is too expensive, define fWert by measuring only A using 1. c) If you want to rely on random decisions, set fWert constant 3. Identify rules that have the same fWert-value and design fselect as tiebreaker (random decisions are best expressed using fselect) 08/04/2010 CPSC 433 Artificial Intelligence 13 Search Paradigms Concrete Example (2): Resolution • We describe our world by a collection of special logical formulas, so-called clauses: L1(t11,…,t1,n1) … Lm(tm1,…,tm,nm) where Li is a predicate symbol or its negation, tij terms out of function symbols and variables (x,y…) variables in different clauses are disjunct • Examples: pq, P(a,b,x)R(x,y,c), Q(f(a,b),g(x,y)), Q(a,b) • A consequence we want to prove is negated, transformed into clauses and these clauses are added to the world. • The consequence is proven, if the empty clause () can be deduced. 08/04/2010 CPSC 433 Artificial Intelligence 14 Search Paradigms Concrete Example (2): Resolution • We derive new clauses by either Resolution or Factorization Resolution: Most General Unifier C P , D P’ (C D) if = mgu(P,P’) Factorization: C P P' (C P) if = mgu(P,P’) Note: clauses commute: pq qp, pqr same as rpq 08/04/2010 CPSC 433 Artificial Intelligence 15 Search Paradigms Concrete Example (2): Resolution Needed: Unification to compute mgu Yet another set-based search problem: States: set of term equations u v, with indicating failure Extension rules: Delete: Decompose: Orient: E {t t} E {f(t1,…,tn) f(s1,…,sn)} E {t x} E E {t1 s1,…,tn sn} E {x t} if t no var. 08/04/2010 CPSC 433 Artificial Intelligence 16 Search Paradigms Concrete Example (2): Resolution Substitute: if x not in t, with s[xt] the E {x t, t' s'} term generated out of s by E {x t, t'[xt] s'[xt]} exchanging all occurences of x by t Occurcheck: E {x t} if x in t Clash: E {f(t1,…,tn) g(s1,…,sn)} if f g Goal condition: all equations in the state have form x t and Occurcheck and Substitute are not applicable 08/04/2010 CPSC 433 Artificial Intelligence 17 Search Paradigms Concrete Example (2): Resolution x,y,z variables Examples for Unification: (1)f(g(x,y),c) f(g(f(d,x),z),c) (2)h(c,d,g(x,y)) h(c,d,g(g(a,y),z)) Examples for Resolution: (1)p q, p q, p q, p q (2)P(x) R(x), R(f(a,b)), P(g(a,b)) (3)P(x) R(y), R(f(a,b)), P(g(a,b)) 08/04/2010 CPSC 433 Artificial Intelligence 18 Search Paradigms Concrete Example (2): Resolution Tasks: • Describe Resolution as set-based search model • Given the following control idea, describe formally a search control for your model, so that we have a search process: Perform factorization whenever possible; choose the smallest possible clauses for resolution; if several clause pairs are smallest, use an ordering <Lit on the predicates and terms 08/04/2010 CPSC 433 Artificial Intelligence 19 Search Paradigms Concrete Example (2): Resolution Tasks (cont.): • Apply your process to the search instance to the following set of clauses: { P(x,y) P(y,x), P(f(x),g(y)) R(y), P(g(x),f(x)), R(x) Q(x,b), Q(a,x)} 08/04/2010 CPSC 433 Artificial Intelligence 20 Search Paradigms Concrete Example(3) Blocks World We want to build an arch in 6x4 coordinate world containing 3 3x1 blocks. c C B a b A S0 = {(A,h,2,1), (B,h,2,2), (C,h,2,3)} G = {{,,} | a,b,c =(a,v,?,?) /\ =(b,v,?,?) /\ =(c,h,?,?) /\ on (c,a) /\ on (c,b) /\ ~touch(a,b) } 08/04/2010 CPSC 433 Artificial Intelligence 21 Search Paradigms Concrete Example(3) Blocks World F: (name, orientation, x-coord, y-coord) covers(a,x,y) = (a,v,x,y) \/ (a,v,x,y-1) \/ (a,v,x,y-2) //vertical case \/ (a,h,x,y) \/ (a,h,x-1,y) \/ (a,h,x-2,y) //horizontal case overlaps(a,b) = x,y covers(a,?,x,y,) /\ covers(b,?,x,y) on(b,a) = (a,v,x,y) /\ covers(b,x,y+4) \/ //vertical (a,h,x,y) /\ (covers(b,x,y+1) \/ covers(b,x+1,y+1) \/ covers(b,x+2,y+1)) //horz touch(a,b) = (a,h,x,y) /\ covers(b,x+4,y) \/ //horizontal (a,v,x,y) /\ (covers(b,x+1,y) \/ covers(b,x+1,y+1) \/ covers(b,x+1,y+2)) \/ //v on(a,b) \/ touch(b,a) //up/down and reverse case 08/04/2010 CPSC 433 Artificial Intelligence 22 Search Paradigms Concrete Example(3) Blocks World Env = a,x,y ((a,v,x,y) -> x<7 /\ y<3) /\ ((a,h,x,y) -> x<5 /\ y<5) //in the borders /\ (b | b≠a ~overlaps(a,b)) //no overlap /\ ((b on(a,b)) \/ (a,?,?,1)) //gravity S = {{(A,o,x,y),(B,o’,x’,y’),(C,o’’,x’’,y’’)} | Env} //type of state is a set of A,B,C blocks Ext = {ff’ | s:S | f s (s-ff’) S} //can move a block anywhere legal fwert = 10 + ((a,b (a,v,?,?) /\ (b,v,?,?) /\ touch(a,b)) ? -5 : 0) + //vert’s shouldn’t touch ((a,b (a,h,?,?) /\ (b,h,?,?) /\ on(a,b)) ? -2 : 0) //don’t stack horizontals fselect = <random> Tset = {(s,s')| AB Ext A s s'=(s-A)B} //from the basic def. of Tset 08/04/2010 CPSC 433 Artificial Intelligence 23 Search Paradigms Concrete Example(3) Blocks World move(C,v,1,1) move(C,v,5,1) move(C,v,6,1) move(C,v,5,1) move(B,v,5,1) move(B,v,6,1) move(B,v,1,1) move(B,v,6,1) move(C,v,5,1) move(B,v,1,1) move(B,v,5,1) 8 10 10 10 5 8 10 5 fwert fwert fwert fselect fwert move(A,v,2,1) move(A,v,3,1) move(A,v,4,1) 5 10 5 fwert fwert move(B,h,1,4) 10 G! 08/04/2010 CPSC 433 Artificial Intelligence 24 Search Paradigms Remarks • Set-based search states can very quickly get very large. • Usually a lot of extensions are possible control is very important • Almost all evolutionary search approaches are set-based (see labs and later example) 08/04/2010 CPSC 433 Artificial Intelligence 25 Search Paradigms