1 / 52

Automating Common Sense Reasoning

Automating Common Sense Reasoning. Gopal Gupta Elmer Salazar, Kyle Marple, Zhuo Chen, Farhad Shakerin (and many others) Department of Computer Science The University of Texas at Dallas Support from NSF is gratefully acknowledged. Artificial Intelligence.

stetson
Download Presentation

Automating Common Sense Reasoning

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Automating Common Sense Reasoning Gopal Gupta Elmer Salazar, Kyle Marple, Zhuo Chen, Farhad Shakerin (and many others) Department of Computer Science The University of Texas at Dallas Support from NSF is gratefully acknowledged

  2. Artificial Intelligence • Intelligent reasoning by computers has been a goal of computer scientists ever since computers were first invented in the 1950s. • Intelligence has two broad components: • Acquiring knowledge (machine learning) • Applying knowledge that is known (automated reasoning) + Vision, Speech Processing, Speech Generation, etc., etc. • Our focus: automated reasoning • Reasoning is essential: machine learning algorithms learn rules that have to be employed for reasoning

  3. Automated Reasoning • Automated reasoning has had its ups and down due to the difficulties involved in building deduction systems based on classical logic • The current AI excitement is driven by machine learning • To build effective AI systems, reasoning methods have to be simplified • Humans have simplified the burden of reasoning by using “common sense reasoning” • Use of probabilistic reasoning & fuzzy sets can be seen as other ways of simplifying reasoning • Classical logic as the basis for common sense reasoning has not worked due to its undecidability, incompleteness & monotonicity

  4. Commonsense Reasoning (CSR) • Humans simplify the burden of reasoning by using • Defaults: E.g., Normally birds fly • Exceptions: penguins, ostriches, wounded birds • Commonsense reasoning also requires: • Nonmonotonicity: revise earlier conclusion in light of new information • allows us to jump to conclusions, but if contradictory information is discovered later, things don’t break down (as in classical logic) • jumping to conclusion == drawing conclusions in absence of info: • Can’t tell if it is cold outside. If I see no one wearing a jacket, it must be warm • You text your friend in the morning; He does not respond; normally he responds right away. You may conclude: he must be taking a shower. • Requirement: be able to reason with negation as failure

  5. Classical Negation vs Negation as Failure • Classical negation • represented as –p • An explicit proof of falsehood of predicate p is needed • -robbed(sutton, bank_one) holds true only if there is an explicit proof of Willie Sutton not robbing Bank One • Negation as failure (NAF) • represented as not(p) • We try to prove proposition p,if we fail, we conclude not(p) is true • No evidence of p then conclude not(p) • not(robbed(sutton, bank_one)) holds true if we fail to find any proof that Willie Sutton robbed Bank_One • Answer Set Prog: paradigm that includes both classical negation and negation as failure

  6. Answer Set Programming • Popular formalism for non monotonic reasoning • Rules of the form: p :-a1 ,…, am, not b1,…, not bn. m≥0, n≥0 (rule) p. (fact) • Another reading: add p to the answer set if a1 ,…, am are in the answer set and b1 ,…, bn are not • Applications to common sense reasoning, planning, constrained optimization, etc. • Semantics given via lfp of a residual program obtained after “Gelfond-Lifschitz” transform • Popular implementations: Smodels, DLV, CLASP, etc. • Almost 30 years of research invested

  7. Negation as Failure • Humans use negation as failure all the time. • Hard (for humans) to deal with nested negation though • Who all will go to Mexico? • Code this as: p :- not s. s :- not r. r :- not p. r :- not s. • What is the semantics of this program? • Individual rules easy to understand; extremely hard to understand what the program means as a whole • Paul will go to Mexico if Sally will not go to Mexico • Sally will go to Mexico if Rob will not go to Mexico. • Rob will go to Mexico if Paul will not go to Mexico. • Rob will go to Mexico if Sally will not go to Mexico.

  8. Answer Set Programming • Answer set programming (ASP) • Incorporates circular reasoning • Based on Possible Worlds and Stable Model Semantics • Given an answer set program, find its models • Model: assignment of true/false value to propositions to make all formulas true. Models are called answer sets • Captures default reasoning, non-monotonic reasoning, constrained optimization, exceptions, weak exceptions, preferences, etc., in a natural way • “Better way to build automated reasoning systems” • Caveats • p ⇐ a, b. really is taken to be p ⇔ a, b. • We are only interested in supported models: if p is in the answer set, it must be in the LHS of a ‘successful’ rule

  9. ASP • Given an answer set program, we want to find out the knowledge (propositions) that it entails (answer sets) • For example, given the program: p :- not q. q :- not p. the possible answer sets are: 1. {p} i.e., p = true, q = false 22. {q} i.e, q = true, p = false • Computed via Gelfond-Lifschitz Method (Guess & Check) • Given an answer set S, for each p  S, delete all rules whose body contains “not p”; • delete all goals of the form “not q” in remaining rules • Compute the least fix point, L, of the residual program • If S = L, then S is an answer set p = Tom teaches DB q = Mary teaches DB Two worlds: Tom teaches DB, Mary does not Mary teaches DB, Tom does not

  10. Finding the Answer Set • Consider the program: • p :- not q. t. r :- t, s. • q :- not p, r. s. h :- p, not h. • Is {p, r, t, s} the answer set? • Apply the GL Method -- If x in answer set, delete all rules with not x in body -- Next, remove all negated goals from the remaining program {p, r, t, s, h} -- Find the LFP of the program: -- Initial guess {p, r, t, s} ≠ {p, r, t, s, h} so {p, r, t, s} is not a stable model.

  11. Finding the Answer Set • Consider the program: • p :- not q. t. r :- t, s. • q :- not p, r. s. h :- p, not h. r. • Is {q, r, t, s} the answer set? • Apply the GL Method -- If x in answer set, delete all rules with not x in body -- Next, remove all negated goals from the remaining program {q, r, t, s} -- Find the LFP of the program: -- Initial guess {q, r, t, s} = LFP so {q, r, t, s} is a stable model.

  12. ASP: Example • Consider the following program, A: p :- not q. t. r :- t, s. q :- not p. s. A has 2 answer sets: {p, r, t, s} & {q, r, t, s}. • Now suppose we add the following rule to A: h :- p, not h. (falsify p) Only one answer set remains: {q, r, t, s} • Consider another program: p :- not s. s :- not r. r :- not p. r :- not s. What are the answer sets? h :- p, not h. • Paul will go to MX if Sally will not go to MX • Sally will go to MX if Rob will not go to MX. • Rob will go to MX if Paul will not go to MX. • Rob will go to MX if Sally will not go to MX. {p, r}

  13. Constraints • The rules that cause problem are of the form: h :- p, q, r, not h. that implicitly declare p to be false • ASP also allows rules of the form: :- p, q, r. which asserts that the conjunction of p, q, and r should be false. • The two are equivalent, except that in the first case, not h may be called indirectly: h :- p, q, s. s :- r, not h. • Constraints are responsible for nonmonotonic behavior • A rule of the form p :- not p wrecks the whole knowledge base

  14. Slide from Son Tran of NMSU

  15. Slide from Son Tran of NMSU

  16. Slide from Chita Baral (ASU) Defaults and Exceptions AS = {flies(tweety), flies(sam)} AS = {flies(tweety))} flies(sam) does not hold any more Our reasoning is aggressive: if we don’t know anything about a bird, it can fly

  17. Slide from Chita Baral (ASU) Closed World Assumption • CWA: if we fail to prove, then take it as a definite • proof of falsehood • We humans use CWA all the time • Make it explicit in ASP via the use of classical negation

  18. Slide from Chita Baral (ASU) Open World Assumption • OWA == No CWA; But now we can be more selective: • CWA for some and OWA for others. CWA about bird, penguin, ab OWA about flies

  19. Slide from Chita Baral (ASU) Open World Assumption • Next, remove CWA about bird, penguin, ab (assume our • information about these concepts is incomplete) But, now that we no longer have CWA about being a penguin, it’s possible that et might be a penguin. We may choose to be more conservative in our reasoning: declare et abnormal only if we fail to prove that et is not a penguin for sure, i.e. X flies if X is a bird and not a penguin for sure (-penguin(X)) flies(et) will now fail, as I fail to establish that et is not a penguin Does et fly? Answer is YES Does et fly now?: Answer is NO

  20. Slide from C. Baral (ASU) University Admission To put all this in perspective, consider the college admission process: Since we have no information about John being special or –special, both eligible(john) and –eligible(john) fail. So John will have to be interviewed

  21. Incomplete Information • Consider the course database: • By default professor P does not teach course C, if teach(P,C) is absent. • The exceptions are courses labeled “staff”. Thus: ¬teach(P,C) :- not teach(P,C), not ab(P,C). ab(P,C) :- teach(staff, C). • Queries ?- teach(mike, pl) and ?- ¬teach(mike,pl) will both fail: we really don’t know if mike teaches pl or not Represented as: teach(mike,ai). teach(sam,db). teach(staff,pl).

  22. Combinatorial Problems: Coloring Slide from S. Tran (NMSU)

  23. Combinatorial Problems: Coloring Slide from S. Tran (NMSU)

  24. Current ASP Systems • Very sophisticated and efficient ASP systems have been developed based on SAT solvers: • CLASP, DLV, Cmodels, Smodels • These systems work as follows: • Ground the programs w.r.t. the constants present in the program • Transform the propositional answer set programs into propositional formulas and then find their models using a SAT solver • The models found are the stable models of the original program • Because SAT solvers require formulas to be propositional, programs with only constants and variables are feasible • Problem: Current systems limited to programs that are finitely groundable (lists and structures are not allowed)

  25. Current ASP Systems: Issues • Problem 1: Program has to be finitely groundable • Not possible to have lists, structures, and complex data structures • Not possible to have arithmetic over reals • Problem 2: Grounding can result in exponential blowup • Given the clause: p(X, a) :- q(Y, b, c). • It turns into 3 x 3, i.e., 9 clauses p(a, a) :- q(a, b, c). p(a, a) :- q(b, b, c). p(a, a) :- q(c, b, c). p(b, a) :- q(a, b, c). p(b, a) :- q(b, b, c). p(b, a) :- q(c, b, c). p(c, a) :- q(a, b, c). p(c, a) :- q(b, b, c). p(c, a) :- q(c, b, c). • Imagine a large knowledgebase with1000 clauses with 100 variables and 100 constants; • Programmers have to contort themselves while writing ASP code • Programs cannot contain lists structures and complex data structures: result in infinite-sized grounded program

  26. Current ASP Systems: Issues • Problem 3: SAT solvers find the entire model • Entire model may contain lot of unnecessary information • I want to know the path from Boston to NYC, the model will contain all possible paths from every city to every other city (overkill) • Problem 4: Some times it may not even be possible to find the answer sought, as they are hidden in the answer set • Answer set of Tower of Hanoi program contains large # of moves • The moves that constitute the actual answer hard to identify • Problem 5: Minor inconsistency in the answer set will result in the system declaring that there are is no answer set • We want to compute an answer if it steers clear of the inconsistent part of the knowledge base • Impossible to have a large knowledge base that is 100% consistent

  27. Why Goal-directed ASP? • Most of the time, given a theory, we are interested in knowing if a particular goal is true or not. • Top down goal-directed execution provides operational semantics (important for usability) • Most practical examples add a constraint to force the answer set to contain a certain goal. • E.g. Zebra puzzle: :- not satisfied. • Answer sets of non-finitely groundable programs computable • Constraints (CLP(R), CLP(FD)) incorporated in the style of current LP systems • Why check the consistency of the whole KB? • Inconsistency in some unrelated part will scuttle us

  28. Solution • Develop goal-directed answer set programming systems that support predicates • Goal-directed means that a query is given, and a proof for the query found by exploring the program search space • Essentially, we need Prolog style execution that supports stable model semantics-based negation • Thus, part of the knowledge base that is only relevant to the query is explored • Predicate answer set programs are directly executed without any grounding: lists and structures also supported Realized in the s(ASP) system developed at UT Dallas Predecessor: Galliwasp (top-down propositional ASP)

  29. s(ASP) System • s(ASP) • Prolog extended with negation based on stable-model semantics • general predicates allowed; • goal-directed execution: user writes ASP code, issues queries • With s(ASP): • Problem #1 & #2 (grounding and explosion) are irrelevant as there is no grounding • Problem #3: only the partial model to answer the query is computed • Problem #4: answer is easily discernible due to query driven nature • Problem #5: consistency checks can be performed incrementally so that if the knowledge base is inconsistent, consistent part of the knowledge base is still useful • Available on sourceforge; includes justification & abduction • s(CASP): incorporates constraints, tabling, DCC

  30. Goal-directed execution of ASP • Key concept for realizing goal-directed execution of ASP: • coinductive execution • Coinduction: dual of induction • computes elements of the GFP • In the process of proving p, if p appears again, then p succeeds • Given: there are two possibilities: • Both jack and jill eat (GFP) • Neither one eats (LFP) eats_food(jack) :- eats_food(jill). eats_food(jill) :- eats_food(jack). ?- eats_food(jack) eats_food(jill). eats_food(jack) coinductive success coinductive hypothesis set = {eats_food(jack), eats_food(jill)}

  31. Goal-directed execution of ASP • To incorporate negation in coinductive reasoning, we need a negative coinductive hypothesis rule: • In the process of establishing not(p), if not(p) is seen again in the resolvent, then not(p) succeeds • In the process of establishing p, if not(p) is seen, computation fails & backtracking occurs • Also, not not p reduces to p.

  32. Goal-directed execution of ASP • Distincition between even loops and odd loops: • Even loop: p recursive calls itself with even number of intervening negations between the two • Odd loop: ….. odd number …… (OLON rules; constraints) • Intuition: even loops generate worlds, odd loops kill worlds • Even-loops succeed by coinduction p :- not q. q :- not p. ?- p  not q  not not p  p (coinductive success) The coinductive hypothesis set is the answer set: {p, not q} • For each constraint (OLON) rule, we extend the query • Given a query Q for a program that contains a constraint rule p :- q, not p. (q ought to be false) extend the query to: ?- Q, (p ∨ not q)

  33. Goal-directed ASP • Consider the following program, A: p :- not q. t. r :- t, s. q :- not p. s. A has 2 answer sets: {p, r, t, s} & {q, r, t, s}. • Now suppose we add the following rule to A: h :- p, not h. (falsify p) • Only one answer set remains: {q, r, t, s} • Recall: • Even loops over negation produce possible worlds (generate) • Constraints or odd loops over negation remove them (test)

  34. Goal-directed ASP • Consider the following program, A’: p :- not q. t. r :- t, s. q :- not p, r. s. h :- p, not h. • Separate into constraint and non-constraint rules: only 1 constraint rule in this case. • Execute the query under co-LP, candidate answer sets will be generated. • Keep the ones not rejected by the constraints. • Suppose the query is ?- q. Execution: q  not p, r  not not q, r  q, r  r  t, s  s  success. Ans = {q, r, t, s} • Next, we need to check that constraint rules will not reject the generated answer set: (not p ; h) • (it doesn’t in this case)

  35. Augmenting the Query • In general, for the constraint rules of p as head, p1:- B1. p2:- B2. ... pn :- Bn., generate rule(s) of the form: chk_p1 :- not(p1), B1. chk_p2 :- not(p2), B2. ... chk_pn :- not(p), Bn. • Generate: nmr_chk :- not(chk_p1), ... , not(chk_pn). • For each pred. definition, generate the dual: not_p :- not(B1), not(B2), ... , not(Bn). • If you want to ask query Q, then ask ?- Q, nmr_chk. • Execution keeps track of atoms in the answer set (PCHS) and atoms not in the answer set (NCHS). not(chk_p1) == p1 \/ not B1

  36. Goal-directed ASP • Consider the following program, P1: (i) p :- not q. (ii) q:- not r. (iii) r :- not p. (iv) q :- not p. P1 has 1 answer set: {q, r}. • Separate into: 3 constraint rules (i, ii, iii) 2 non-constraint rules (i, iv). p :- not(q).q :- not(r).r :- not(p).q :- not(p). chk_p :- not(p), not(q).chk_q :- not(q), not(r).chk_r :- not(r), not(p). nmr_chk :- not(chk_p), not(chk_q), not(chk_r). not_p :- q. not_q :- r, p.not_r :- p. Suppose the query is ?- r. Expand as in co-LP: r  not p  not not q  q (  not r  fail, backtrack)  not p  success. Ans={r, q} which satisfies the constraint rules of nmr_chk.

  37. Implementing Predicate ASP Systems • Aside from coinduction many other innovations needed to realize the s(ASP) system: • Dual rules to handle negation • Constructive negation support (domains are infinite) • Universally Quantified Vars (due to negation & duals) • A special infinite Herbrand Universe • s(ASP): Essentially, Prolog with stable model-based negation • Has been used for implementing non-trivial applications: • Check if an undergraduate student can graduate at a US university • Physician advisory system to manage chronic heart failure • Representing cell biology knowledge as an ASP program

  38. Predicate ASP Systems: Challenges • Constructive negation: • For each variable carry the values it cannot be bound to X \= a unifies with Y \= b: result X \= {a, b} Disunification of such variables is complex (not allowed) How to detect failure when a variable is constrained against every element of the universe? • Dual rules to handle negation • Need to handle existentially quantified variables p(X) :- q(X, Y). • Programs such as: p(X) :- not q(X). q(X) :- not p(X). produce uncountable number of models Computing Stable Models of Normal Logic Programs without Grounding. Marple, Salazar, Gupta, UT Dallas Tech Rep. s(ASP) available on sourceforge

  39. Dual Rules • Negation is handled using dual rules • The propositional case is simple: given the rules: p :- not q. p :- r, not s. • Rule for not p is computed (completion) not p :- np1, np2. np1 :- q. np2 :- not r. np2 :- s. • Whenever not p is called, dual rule executed • Each goal gets its own dual clause p :- not q; (r, not s)

  40. Dual Rules (cont’d) • Predicates complicate things because they can create bindings; Given the rules: • p(X) :- not q(X). • p(X) :- r(X), not s(X). • The call r(X) may modify X before not s(X) is called • The dual must account for this; the solution is to include the previous goals when computing the dual clause for a given goal: • not p(X) :- np1(X), np2(X). • np1(X) :- q(X). • np2(X) :- not r(X). • np2(X) :- r(X), s(X). Nice innovation by Kyle & Elmer (tabling will help with efficiency, hence)

  41. Universally Quantified Vars • Variables in a clause also have implicit quantifiers • Dual rules must handle these correctly. The rule: • p(X) :- not q(X, Y). • is equivalent to the formula: • (p(X) ←Y ¬q(X,Y )) • The dual of this formula is: • (¬p(X) ←Y q(X,Y )) • Thus, the dual clause has a universally quantified var in its body • Problem: all “body variables” are existentially quantified E XA XA A

  42. Forall Mechanism • Our solution was to create a “forall” mechanism • Given the rule: • p(X) :- not q(X, Y). • the dual will be: • not p(X) :- forall(Y, np1(X, Y)). • np1(X, Y) :- q(X, Y). • To execute the forall, the goal np1 is called normally and the forall variable examined: • If the variable is unbound, the forall succeeds • If the variable is bound, failure and backtracking occur • If the variable is negatively constrained, the goal is executed for each constrained value, substituting the value for the variable • The forall will only succeed if the goal succeeds for all values of the variable

  43. Special s(ASP) Universe • What would happen if a variable were constrained against every element of its domain? • To prevent this situation, s(ASP) uses a special Herbrand Universe (HU) that is an infinite superset of the HU • There will always be elements in the s(ASP) universe against which a variable cannot be constrained • These elements are purely theoretical • If a variable is now constrained against every element of the Herbrand Universe, s(ASP) can succeed using these elements • These new solutions will never be produced (and are not inconsistent with the result) from a grounded program • They represent different, equally correct information

  44. s(ASP) Universe AS = {d(1), not p(1)}

  45. Applications Developed with s(ASP) • Grad audit system: figure out if an undergraduate student at a university can graduate • Complex requirements expressed as ASP rules • 100s of courses that students could potentially choose from • Lists and structures come in handy in keeping the program compact • Developed over a few weeks by 2 undergraduates • Catalog changes frequently: ASP rules easy to update

  46. Applications Developed • Physician advisory system: advises doctors for managing chronic heart failure in patients • Automates American College of Cardiology guidelines • Complex guidelines expressed as rules (60 odd rules in 80 pages) • Knowledge patterns developed to facilitate the modeling • Tested with UT Southwestern medical center; • found diagnosis that was missed by the doctor • Representing high school cell biology knowledge: • represented using ASP • Questions can be posed as ASP queries to the system • Natural lang. Q&A systems (make use of defaults) • Recommendation systems – birthday gift advisor

  47. Recommendation Systems • Recommendation systems in reality need sophisticated modeling of human behavior • Current systems based on machine learning, and are not too precise, in our opinion • Birthday gift advisor: • model the generosity, wealth level of the person • model the interests of the person receiving the gift • use the information to determine the appropriate gift for one of your friends • very complex dynamics go in determining what gift you will get a person • In general, any type of recommendation system can be similarly built.

  48. Intelligent IoT • Intelligent systems that take as input sensor signals and determine the action based on some logic can be developed • The situation where we take an action in the absence of a signal arises frequently • Example: intelligent lighting system turn_on_light:- motion_detected, door_moved. turn_on_light:- light_is_on, motion_detected. turn_on_light:- light_is_on, not(motion_detected), not(door_moved). -turn_on_light:- not(motion_detected), door_moved.

  49. s(ASP) Hackathon • 2 weeks of teaching ASP; 150 signed up; 18 projects

  50. Conclusions • ASP: a comprehensive methodology for developing intelligent applications; common sense reasoning easily simulated • Incorporates both circular and well-founded reasoning • SAT-based implementations are fast, but face problems wrt building general purpose, large scale applications • s(ASP) adds stable model-based negation to Prolog, but now inherits all of Prolog’s problems (inefficient search mechanism) • Challenges: make efficiency competitive with SAT-based impl. • Integrate CLP(FD) • Integrate CLP(R) • Add tabled execution • Design an abstract machine; efficient (or-parallel) implementation • Develop more large-scale applications Work by visiting PhD student Joaquin Arias from IMDEA, Spain. Available as s(CASP) system on github

More Related