1. Ontologies Reasoning Components Agents Simulations Reasoning with Classical Propositional Logic Jacques Robin

2. Outline  Syntax  Full CPL  Implicative Normal Form CPL (INFCPL)  Horn CPL (HCPL)  Semantics  Cognitive and Herbrand interpretations, models  Reasoning  FCPL Reasoning  Truth-tabel based model checking  Multiple inference rules  INFCPL Reasoning  Resolution and factoring  DPLL  WalkSat  HCPL Reasoning  Forward chaining  Backward chaining

3. Full Classical Propositional Logic (FCPL): syntax FCPLFormula Syntax FCPLUnaryConnective Connective: enum{  } FCPLBinaryConnective Connective: enum{ , , ,  } FCPLConnective Functor ConstantSymbol Arg 1..2  (a  (b  ((  c  d)  a)  b))

4. CPL Normal Forms ConstantSymbol INFCPLFormula Functor =  INFCPLClause Functor =  INFCLPLHS Functor =  INFCLPRHS Functor =  Premisse Conclusion Implicative Normal Form (INF) Conjunctive Normal Form (CNF) CNFCPLFormula Functor =  CNFCPLClause Functor =  NegativeLiteral Functor =  Literal ConstantSymbol Semantic equivalence: a  b  c  d   (a  b)  c  d   a   b  c  d * * * **

5. Horn CLP ConstantSymbol INFCPLFormula Functor =  INFCPLClause Functor =  INFCLPLHS Functor =  INFCLPRHS Functor =  Premisse Conclusion Implicative Normal Form (INF) Conjunctive Normal Form (CNF) CNFCPLFormula Functor =  CNFCPLClause Functor =  NegativeLiteral Functor =  LiteralConstantSymbol * * * * DefiniteClause IntegrityConstraint Fact DefiniteClause IntegrityConstraint Fact context IntegrityConstraint inv IC: Conclusion.ConstantSymbol = false context DefiniteClause inv DC: Conclusion.ConstantSymbol  false context Fact inv Fact: Premisse -> size() = 1 and Premisse -> ConstantSymbol = true context IntegrityConstraint inv IC: Literal->forAll(oclIsKindOf(NegativeLiteral)) context DefiniteClause inv DC: Literal.oclIsKindOf(ConstantSymbol)->size() = 1 context Fact inv Fact: Literal->forAll(oclIsKindOf(ConstantSymbol)) a  b  c  false a  b  c  d true  d  a   b   c  a   b   c  d d

6. FCPL semantics: cognitive interpretation FCPLFormulaConstantSymbol FCPLConnective FCPLUnaryConnective Connective: enum{  } FCPLBinaryConnective Connective: enum{ , , ,  } TruthValue Value: enum{true,false} FCLPCognitiveInterpretation Syntax Semantics Arg Functor CompoundDomainProperty FormulaMapping AtomicDomainProperty ConstantMapping 1..2  (a  (b  ((  c  d)  a)  b)) csm1(pitIn12) = agent knows there is a pit in coordinates (1,2) csm2(pitIn12) = John is the King of England fm1(pitIn12   pitIn11) = agent knows there is a pit in coordinates (1,2) and no pit in coordinates (1,1) fm1(pitIn12   pitIn11) = John is the Kind of England and John is not the King of France

7. FCPL semantics: Herbrand interpretation FCPLFormulaConstantSymbol FCPLConnective FCPLUnaryConnective Connective: enum{  } FCPLBinaryConnective Connective: enum{ , , ,  } TruthValue Value: enum{true,false} FCLPHerbrandInterpretation Syntax Semantics Arg Functor ConstantValuation FormulaValuation FCLPHerbrandModel 1..2  (a  (b  ((  c  d)  a)  b)) {cv1(pitIn12) = true, cv1(pitIn11) = true,...} {cv2(pitIn12) = true, cv2(pitIn11) = false,...} {fv1(pitIn12   pitIn11) = true, fv1(pitIn12  pitIn11) = true,...} {fv2(pitIn12   pitIn11) = true, fv2(pitIn12  pitIn11) = false,...}

8. FCPL semantics FCPLFormulaConstantSymbol FCPLConnective CompoundDomainProperty FCPLUnaryConnective Connective: enum{  } FCPLBinaryConnective Connective: enum{ , , ,  } AtomicDomainProperty TruthValue Value: enum{true,false} FCLPCognitiveInterpretation FCLPHerbrandInterpretation Syntax Semantics Arg Functor ConstantValuationFormulaValuation FCLPHerbrandModel FormulaMapping ConstantMapping 1..2  (a  (b  ((  c  d)  a)  b))

9. Entailment and models  Entailment |=:  f |= f’ iff:  H i, H i (f) = true  H i (f’) = true  Logical equivalence  :  f  f’ iff f |= f’ and f’ |= f  Herbrand model:  An Herbrand interpretation H i is a (Herbrand) model of formula f iff its truth value corresponds to the application of the truth-table definition of the FCPL connectives to the truth value in H i of the constant symbols that compose f  f valid (or tautology) iff true in all H i (f), ex, a   a  f satisfiable iff true in at least one H i (f)  f unsatisfiable (or contradiction) iff false in all H i (f), ex, a   a Valid formulas Satisfiable formulas Unsatisfiable formulas

10. Logic-Based Agent Ask Tell Retract Environment Actuators Knowledge Base B: Domain Model in Logic L Inference Engine: Theorem Prover for Logic L Given B as axiom, formula f is a theorem of L? B |= L f ? B  f is valid in L? (Boolean CSP search proof) B   f is unsatisfiable in L? (Refutation proof) Sensors  Strenghts:  Reuse results and insights about correct reasoning that matured over 23 centuries  Semantics (meaning) of a knowledge base can be represented formally as syntax, a key step towards automating reasoning

11. Truth-table based model checking  To answer: Ask(  )  Enumerate all H i s from domain proposition alphabet  Use truth-table to compute M h (KB) and M h (  )  If M h (KB)  M h (  ), then answer yes, else answer no  Example:  KB =  pit11   breeze11   pit12  breeze12   1 =  pit21   2 =  pit22

12. FCLP inference rules  Bi-directional (logical equivalences) R 1 : f  g  g  f R 2 : f  g  g  f R 3 : (f  g)  h  f  (g  h) R 4 : (f  g)  h  f  (g  h) R 5 :  f  f R 6 : f  g   g   f R 7 : f  g   f  g R 8 : f  g  (f  g)  (g  f) R 9 :  (f  g)   f   g R 10 :  (f  g)   f   g R 11 : f  (g  h)  (f  g)  (f  h) R 12 : f  (g  h)  (f  g)  (f  h) R 13 : f  f  f %factoring  Directed (logical entailments) R 14 : f  g, f |= g %modus ponens R 15 : f  g,  g |=  f %modus tollens R 16 : f  g |= f %and-elimination R 17 : l 1 ...  l i ... l k, m 1 ...  m j-1   l i  m j-1... m k |= l 1 ...  l i-1  l i-1... l k  m 1 ...  m j-1  m j-1... m k %resolution

13. Multiple inference rule application Idea:  KB |= f ?  KB 0 = KB  Apply inference rule: KB i |= g  Update KB i+1 = KB i  g  Iterate until f  KB k or until f  KB n and KB n+1 = KB n  Transforms proving KB |= f into search problem  At each step:  Which inference rule to apply?  To which sub-formula of f? Example proof:  KB 0 =  P 1,1  (B 1,1  P 1,2  P 2,1 )  (B 2,1  P 1,1  P 2,2  P 3,1 )   B 1,1  B 2,1  Query:  (P 1,2  P 2,1 )  Cognitive interpretation:  B X,Y : agent felt breeze in coordinate (X,Y)  P X,Y : agent knows there is a pit in coordinate (X,Y)  Apply R 8 to B 1,1  P 1,2  P 2,1 KB 1 = KB 0  (B 1,1  (P 1,2  P 2,1 ))  ((P 1,2  P 2,1 )  B 1,1 )  Apply R 6 to last sub-formula KB 2 = KB 1  (  B 1,1   (P 1,2  P 2,1 ))  Apply R 14 to  B 1,1 and last sub-formula KB 3 = KB 2   (P 1,2  P 2,1 )

14. Resolution and factoring  Repeated application of only two inference rules:  resolution and factoring  More efficient than using multiple inference rules  search space with far smaller branching factor  Refutation proof:  Derive false from KB   Query  Requires both in normal form (conjunctive or implicative)  Example proof in conjunctive normal form:

15. Resolution strategies  Search heuristics for resolution-based theorem proving  Two heuristic classes:  Choice of clause pair to resolve inside current KB  Choice of literals to resolve inside chosen clause pair  Unit preference:  Prefer pairs with one unit clause (i.e., literals)  Rationale: generates smaller clauses, eliminates much literal choice in pair  Unit resolution: turn preference into requirement  Set of support:  Define small subset of initial clauses as initial “set of support”  At each step:  Only consider clause pairs with one member from current set of support  Add step result to set of support  Efficiency depend on cleverness of initial set of support  Common domain-independent initial set of support: negated query  Beyond efficiency, results in easier to understand, goal-directed proofs  Linear resolution:  At each step only consider pairs (f,g) where f is either:  (a) in KB 0, or  (b) an ancestor of g in the proof tree  Input resolution:  Specialization of linear resolution excluding (b) case  Generates spine-looking proofs trees

16. FCPL theorem proving as boolean CSP exhaustive global backtracking search  Put f = KB   Query in conjunctive normal form  Try to prove it unsatisfiable  Consider each literal in f as a boolean variable  Consider each clause in f as a constraint on these variables  Solve the underlying boolean CSP problem by using:  Exhaustive global backtracking search  of all complete variable assignments  showing none satisfies all constraint in f  Initial state: empty assignment of pre-ordered variables  Search operator:  Tentative assignment of next yet unassigned variable L i (i th literal in f)  Apply truth table definitions to propagate constraints in which L i appears (clauses of f involving L)  If propagation violates one constraint, backtrack on L i  If propagation satisfies all constraints:  iterate on L i+1  if L i was last literal in f, fail, KB   Query satisfiable, and thus KB |  Query

17. FCPL theorem proving as boolean CSP backtracking search: example  Variables = {B 1,1, P 1,2, P 2,1 }  Constraints: {  B 1,1,  P 1,2  B 1,1,  P 2,1  B 1,1,  B 1,1  P 1,2  P 2,1, P 1,2 } V = [?,?,?] C = [?,?,?,?,?] V = [0,?,?] C = [1,?,?,?,?] V = [1,?,?] C = [0,?,?,?,?] V = [0,0,?] C = [1,0,?,?,0] V = [0,1,?] C = [1,0,?,?,1] V = [1,0,?] C = [0,1,?,?,0] V = [1,1,?] C = [0,0,?,?,1] V = [1,1,1] C = [0,0,0,0,1] V = [1,1,0] C = [0,0,1,0,1] V = [1,0,1] C = [0,1,0,0,0] V = [1,0,0] C = [0,1,1,0,0] V = [0,1,1] C = [1,0,0,1,1] V = [0,1,0] C = [1,0,0,0,1] V = [0,0,1] C = [1,0,0,0,0] V = [0,0,0] C = [1,0,0,0,0]

18. DPLL algorithm  General purpose CSP backtracking search very inefficient for proving large CFPL theorems  Davis, Putnam, Logemann & Loveland algorithm (DPPL):  Specialization of CSP backtracking search  Exploiting specificity of CFPL theorem proving recast as CSP search  To apply search completeness preserving heuristics  Concepts:  Pure symbol S: yet unassigned variable positive in all clauses or negated in all clauses  Unit clause C: clause with all but one literal already assigned to false  Heuristics:  Pure symbol heuristic: assign pure symbols first  Unit propagation:  Assign unit clause literals first  Recursively generate new ones  Early termination heuristic:  After assigning L i = true, propagate C j = true  C j | L i  C j (avoiding truth-table look-ups)  Prune sub-tree below any node where  C j | C j = false  Clause learning

19. Satisfiability of formula as boolean CSP heuristic local stochastic search  DPLL is not restricted to proving entailment by proving unsatisfiability  It can also prove satisfiability of a FCPL formula  Many problems in computer science and AI can be recast as a satisfiability problem  Heuristic local stochastic boolean CSP search more space scalable than DPLL for satisfiability  However since it is not exhaustive search, it cannot prove unsatisfiability (and thus entailment), only strongly suspect it  WalkSAT  Initial state: random assignment of pre-ordered variables  Search operator:  Pick a yet unsatisfied clause and one literal in it  Flip the literal assignment  At each step, randomly chose between to picking strategies:  Pick literal which flip results in steepest decrease in number of yet unsatisfied clauses  Random pick

20. Direct x indirect use of search for agent reasoning Agent Decision Problem Domain Specific Agent Decision Problem Search Model: State data structure Successor function Goal function Heuristic function Domain Specific Knowledge Base Model: Logic formulas true  d f  g  h  c... Domain Independent Inference Engine Search Model State data structure Successor function Goal function Heuristic function Domain Independent Search Algorithm Agent Application Developer Reasoning Component Developer

21. Horn CPL reasoning  Practical limitations of FCPL reasoning:  For experts in most application domain (medicine, law, business, design, troubleshooting):  Non-intuitiveness of FCPL formulas for knowledge acquisition  Non-intuitiveness of proofs generated by FCPL algorithms for knowledge validation  Theoretical limitation of FCPL reasoning:  exponential in the size of the KB  Syntactic limitation to Horn clauses overcome both limitations:  KB becomes base of simple rules If p 1 and... and p n then c, with logical semantics p 1 ...  p n  c  Two algorithms are available, rule forward chaining and rule backward chaining, that are:  Intuitive  Sound and complete for HCPL  Linear in the size of the KB  For most application domains, loss of expressiveness can be overcome by addition of new symbols and clauses:  ex, FCPL KB 1 = p  q  c  d has no logical equivalent in HCPL in terms of alphabet {p,q,c,d}  However KB 2 = (p  q  notd  c)  (p  q  notc  d)  (c  notc  false)  (d  notd  false) is an HCPL formula logically equivalent to KB 1

22. Propositional forward chaining  Repeated application of modus ponens until reaching a fixed point  At each step i:  Fire all rules (i.e., Horn clauses with at least one positive and one negative literal) with all premises already in KB i  Add their respective conclusions to KB i+1  Fixed point k reached when KB k = KB k-1  KB k = {f | KB 0 |= f}, i.e., all logical conclusions of KB 0  If f  KB k, then KB 0 |= f, otherwise, KB 0 |  f  Naturally data-driven reasoning:  Guided by fact (axioms) in KB 0  Allows intuitive, direct implementation of reactive agents  Generally inefficient for:  Inefficient for specific entailment query  Cumbersome for deliberative agent implementations  Builds and-or proof graph bottom-up

23. Propositional forward chaining: example

30. Propositional backward chaining  Repeated application of resolution using:  Unit input resolution strategy with negated query as initial set of support  At each step i:  Search KB 0 for clause of the form p 1 ...  p n  g to resolve with clause g popped from the goal stack  If there are several ones, pick one, push p 1 ...  p n on goal stack, and push other ones alternative stack to consider upon backtracking  If there are none, backtrack (i.e., pop alternative stack)  Terminates:  Successfully when goal stack is empty  As failure when goal stack is non empty but alternative stack is  Naturally goal-driven reasoning:  Guided by goal (theorem to prove)  Allows intuitive, direct implementation of deliberative agents  Generally:  Inefficient for deriving all logical conclusions from KB  Cumbersome implementation of reactive agents  Builds and-or proof graph top-down

31. Propositional backward chaining: example Goal Stack Q Alternative Stack 

32. Propositional backward chaining: example Goal Stack P Alternative Stack 

33. Propositional backward chaining: example Goal Stack L M Alternative Stack 

34. Propositional backward chaining: example Goal Stack A P M Alternative Stack A B

35. Propositional backward chaining: example Goal Stack P M Alternative Stack A B

36. Propositional backward chaining: example Goal Stack A B M Alternative Stack 

37. Propositional backward chaining: example Goal Stack M Alternative Stack 

38. Propositional backward chaining: example Goal Stack B L Alternative Stack 

39. Propositional backward chaining: example Goal Stack  Alternative Stack 

40. Propositional backward chaining: example Goal Stack  Alternative Stack 

41. Propositional backward chaining: example Goal Stack  Alternative Stack 

42. Limitations of propositional logic  Ontological:  Cannot represent knowledge intentionally  No concise representation of generic relations (generic in terms of categories, space, time, etc.)  ex, no way to concisely formalize the Wumpus world rule: “at any step during the exploration, the agent perceiving a stench makes him knows that there is a Wumpus in a location adjacent to his”  Propositional logic:  Requires conjunction of 100,000 equivalences to represent this rule for an exploration of at most 1000 steps of a cavern size 10x10  (stench1_1_1  wumpus1_1_2  wumpus1_2_1) ......  (stench1000_1_1  wumpus100_1_2  wumpus1000_2_1) ......  (stench1_10_10  wumpus1_9_10  wumpus1_10_9) ......  (stench1000_10_10  wumpus100_9_10  wumpus1000_9_10)  Epistemological:  Agent always completely confident of its positive or negative beliefs  No explicit representation of ignorance (missing knowledge)  Only way to represent uncertainty is disjunction  Once held, agent belief cannot be questioned by new evidence (ex, from sensors)