From: AAAI-86 Proceedings. Copyright ©1986, AAAI (www.aaai.org). All rights reserved. IMPLEMENTATION OF AND EXPERIMENTS A VARIABLE PRECISION LOGIC INFERENCE Peter WITH SYSTEM* Haddawy Intelligent Systems Group Department of Computer Science University of Illinois Urbana, Illinois ABSTRACT A approximate system capable inferences presented. under Censored of time production performing is are used to represent tion. These both domain and control informaare given a probabilistic semantics a scheme and reasoning is performed using Examples based on Dempster-Shafer theory. show the naturalness of the representation and the flexibility further of research the system. Suggestions Certainty refers to statement, while specificity of detail constraints rules 61801 the degree of a description. the certainty of an varied to conform to sored and degree inference cost Michalski Precision production control in to the a The idea of a logic which presented by This Variable of belief refers constraints and Logic rules be was Winston [1985]. (VPL) used cen- to encode information. in could both The rules domain take the form for are offered. P->D[C read If P then D unless C. part of the rule I INTRODUCTION The unless It is a Sunday afternoon and your car is taking you for a drive. autonomous fully Suddenly Your to do. If your soning technology, reach mine best any better than because course decision, there are extra-logical time and In cost, resource account. certainty, be used any and to flexibly what it would trying reanever you. even a rough is needed reasoning constraints limitations which tradeoff specificity / SCIENCE Whereas to checking in In this unless guess, is is a sys- exclusive-or time is their truth that of rule resources Thus, only is devoted critical to checking The interpreted between given are a lim- situations. logically operator the consequent. of than unlimited of resources symbol the censor. the antecedents, censors the as censor an and the rule possi-> process we weather is good, Sunday and the weather as must be the of inferences can can Go to the park such between to these priority amount conclude the park. constraints. * Th’IS work was supported in part by the Defense Advanced Research Project Agency under grant N00014-K-85-0878, in part by the National Science Foundation under grant NSF DCR 8406801, and in part by the Office of Naval Research under grant N00014-82-K-0186. 238 devoted Sunday cost adjust antecedents. determination a lower the limit. practical the is given ited the best decision The Therefore, values is called exception conditions and as to be false most of the time. hit the to deter- it would and What time ahead to decide itself of producing a given into while the road by current are of action both ble within taken powered chances indecision. capable into 5 seconds car were destroying situation out car has a decision the truck, tem pulls a truck of you. Censors represent such are considered if it is Sunday I will go to the park; is bad, * - * +D [C,vC,v interpreted is bad, and the and if it is I will not go to Formally, P,AP,A is logically that 1 Weather --. as P,AP,A - - - A-( C,vC,v...)--+D P,AP,A ’ ’ - A(C,VC,V...)--D and In order to make the exceptions quantitative, These numerical parameters are associated with each rule, representing the strength of inference when the truth the value value of the censor is unknown. precision of inferences This of the notion is known These values by the following allow the to be varied. paper presents of uncertainty to conform constrained and when values a formalization in censored When the Shafer interval value according produc- to given II FORMALISM time censor is known, the D is computed s(D) = @‘)[l - p(C)Io p(D) = 1 - s(P)s(C)@ limits. AND of the for the conclusion to tion rules and an implementation of an inference system capable of varying the certainty of inferences are restrictions: THEORY When the censor value is unknown, the formulas are Uncertain tem is using Dempster-Shafer theory information and inference performed A fact scheme [Shafer is represented facts. in the VPL a is 19761. assigned a while the p value intervals indicates for A & 1A The s and p values as the minimum and the proposition. proposition values. the interval disjunction formulas for respectively values of The of vals is similar probabilities of uncertainty the or the product or and plausibility support across rules the employed by Dubois rule A using an approach [1984] & Prade -> [1985]. B, prob(A)EISAPA]. it can be prob(B) E [S,S,, I--s,+S,P,]. prob(d3 / A)E[S,P,] th en P, = l-S, it follows this that scheme, represented prob(B)EISAS,, the we have certainty by four values: of the that Now if from a to and shown l-S,,%]. conditionally to combine conclu- to that independent. two in [Ginsberg which To use rule is Q p 7 S, where The Shafer inter- 19841: above logic, but logic representation, instances and discussion for proposi- the system uses a typed VPL The containing to logic How- a semantics for expressions is needed. Rules entire with of the a short-hand domain certainty with form certainty as Vx,y p(B(x)lA(x,y)) preted = Similarly, [s p] is interpreted free variA(x,y) -> [s p] are inter- for listing of x and y. ground problems. no additional an associated only infor- a finite In this propositional ables essentially type the elements of predicate argument. equivalent present presented scheme ever, B(x), with has inference terms are thus l-(ad+bc) representation. enumerates for each 1 bd “_” approximate mation domain = a=S,forPAlC->D l- I-(ad+bc) predicate derived prob(l3~A)~[S,P,] where Th en Suppose similar and argued -- an as expressing are propogated by Ginsberg are used [u b]@[c d] = tional conditional events formula sum separately. Rules are interpreted probabilities. Beliefs that as by applying probabilistic to evidence of the of a conjunction is calculated for multiply in a difference is represented Th e value of facts The Evidence 1- as the 1 - s(P)/3. sions is combined using Dempster’s orthogonal sum rule. This requires the assumption that the can also be thought of ‘unknown’ [0 l]. s by p(A) = The amount A value p(D) = certainty its plausibility. maximum is defined s(D) = s(P)a of rules [s p]. The a proposition, are related $(-IA). on Domain in the form represented by a Shafer interval, value indicates the support for sys- based [s p]. This rules is over the a fact A(x) as Vx p(A(x)) Is PI. p=S;forPAC->lD 7= S, for P -> s=S;forP-> D YD Uncertainty and Expert Systems: AUTOMATED REASONING / 239 III SYSTEM OVERVIEW approximate that The VPL components: the knowledge and base, the consists interface, the unifier, rule-base implemented the bird such as a penguin parser, tion such as dead. Lisp the engine, The system and runs is on The system for is designed incremental The user may assert base type declarations, followed some Following example runs. rules the system from develop- or retract and Next, changing our When may perform may The rule-base sible query analyzer the chain is a rule from a censor. depths determines inference depths time of censor chain required chaining. of times query times. for each for tree with is stored posall A censor in the search A list for each inference time limit system the the possibility our belief is neither bird, leading in the time associated searches censor time limit The inference, with possibly and breadth first is achieved ground instances If none with the goal. query is given search process, certainty satisfy a fact is found, any it tries If both When on the calculation on a value of the stack attempts stack are fail, the the the on a calculation This ties. / SCIENCE The section first is rover jane tweety road-runner)) (animal)) (bird)) (animal)) (bird)) (bird)) (bird)) (is-kiwi (bird)) (is-domestic-turkey (bird)) (animal)) (animal)) (has-broken-wing (bird)) assert ((is-bird evaluated and put on the bottom 8x) => (flies 8x) 1 (is-special-bird 8x) (is-in-unusual-condition ((is-dead 8x) 1.0 1.0 .9 .05) $x) = > (is-in-unusual-condition $x) 1.0 0.0) to the user query. presents intended $x) = > (is-in-unusual-condition two the system’s to simple capabili- highlight the with decrease road-runner)) (is-emu ((is-sick to demonstrate to further the entries I-V EXAMPLES examples (spot (is-penguin (is-sick performing The computation tweety (flies (animal)) system During terminates, (is-bird (is-dead unify put that combines in his ability to fly. Finally, if tweety in an unusual condition nor a special (tweety after to find rules which are corresponds the of [0 11. the search list. exhaustive with for of his death to fly propor- is told information fly. death he is able to fly. (is-ostrich is goal. instructions stack. stages: the of these this (is-special-bird on strategy variables goal, in tweety’s in his ability (is-in-unusual-condition all consistent unifies a value calculations depth The a which If no chaining in two free certainty bird, cannot pred is unlim- search by generating is a dead tweety the will time. depth exhaustive. To unification. searches for which search The of case backward is performed and search depth limited calculation. (bird base to determine chaining performs Inference censors. it, in which data declara- data an optional in the requried is specified, system have chaining a response ited. search with the time maximum guarantee may tweety that the system be a kiwi, type (animal query by predicate domain associated base. A user condi- file shows the input file is a log of After the rules are loaded, that the certainty of or is in an unusual it concludes changes uniform 240 is told which tionately. analysis of followed define new types and predicates, and make Once a rule base is complete, the user an is kind by rules. queries. facts, idea a to be fully rule The it is a special The input tions, 3640. interactive methods. can fly unless of six main the inference analyzer. in Common Symbolics ment. system user inference a bird ((has-broken-wing (is-in-unusual-condition 8x) 0.9 0.06) $x) = > 8x) 1.0 0.0) ((is-penguin 8x) => ((is-ostrich (is-special-bird $x) => (is-special-bird 8x) 1.0 0) ((is-emu $x) = > (is-special-bird 8x) 1.0 0) ((is-kiwi $x) => tf;x) 1.0 0) (is-special-bird ((is-domestic-turkey 0.027635619 <RESULT> Example Tweety Runs is a dead bird --- file shows in time tion > assert Command > ((is-bird ENTER tweety) Command > ((is-dead tweety) of its brakes limit. or assertion the system the 1 1) ENTER Command or make query and run shows With a censor cannot depth seconds elapsed chaining version A log of the of varying the time of 1 or less depth the truth censor approximate on the condi- road. determine and values thus uses of the of the rule. We of system (level (low medium chaining the of censor The input if a car can stop based the effect road-condition more > (flies tweety) <RESULT> an obstacle 1 1) or assertion 0.103377685 for determining sample Command censor rules to avoid or assertion ENTER using [l.OO 1.001 ity of the system to vary the depth chaining in response to time limits. Command ENTER time The next example is the one in the introduction. It shows the abil- $x) 1.0 0) described ENTER elapsed $x) = > (is-special-bird -- seconds $x) 1.0 0) (rating of UNLIMITED (good high)) fair poor)) (substance (gravel ice)) (place (road ground)) time (temp-type [O.OO 0.001 (looks (below-freezing (shiny moderate hot)) rough)) -- reduce certainty in Tweety’s death ENTER Command > (assert ENTER or make ((is-dead tweety) Command query pred of system (speed-distance-ratio .7 .8)) (road-condition > (? (flies tweety)) using censor 0.1013306 chaining seconds depth elapsed (rating)) (brake-condition of UNLIMITED (rating)) (on (substance time place)) (temperature <RESULT> --- [O.OO 0.3Oj suspect ENTER that Tweety Command > (assert ENTER ((is-kiwi Command or make tweety) is a kiwi --query censor 0.10849539 chaining seconds (temp-type)) (road-appearance (looks)) (construction-site 0) (sound-of-pebbles-hitting-underside-of-car assert .3 .5)) ; rules or assertion ( (- speed-distance-ratio depth elapsed of UNLIMITED (can-stop-in-time) time 1 (road-condition high) Tweety ENTER > (assert ENTER is healthy Command or make query Command Command ( (on gravel of system tweety) censor chaining = > (road-condition poor) 1.0 0) ( (temperature or assertion tweety) road) => (road-condition poor) .9 .l) 0 0)) below-freezing) (road-appearance 1 1) shiny) = > (on ice road) .9 .I) or assertion ( (construction-site) > (? (flies tweety)) using 1.0 1.0 .85 .l) and normal -- (( is - in - unusual-condition > ((I is-special-bird ENTER poor) [O.OO 0.211 ( (on ice road) --- => poor) (brake-condition <RESULT> 0) of system > (? (flies tweety)) using (level)) (can-stop-in-time()) or assertion depth of UNLIMITED Uncertainty (sound-of-nebbles-hitting-underside-of-car1 and Expert Systems: AUTOMATED REASONING / 24 1 => (on gravel road) .9 .l) systems for operating a rule’s censor rooms In the current ; facts ((speed-distance-ratio ((temperature .05 .15) high) .2 .3) below-freezing) ((road-appearance shiny) ((construction-site) values 0 0) system, is known, are used. to domestic if the value the a! and If the value robots. is unknown, the and 6 values are used. A better approach be to look at the degree to which the 1.0 1.0) ((sound-of-pebbles-hitting-underside-of-car) value .8 .85) is known between the and use this to 7 would censor interpolate 7 & S rule o & p and of /3 certainty certainty values. Example Runs --- time limit of 1 second -- expressed ENTER nomies Command or input > (? ((can-stop-in-time) using censor chaining 0.08085977 seconds <RESULT > file 1)) depth of 2 elapsed carry I am working > (? ((can-stop-in-time) using censor 0.031729784 The or make query of system trade-offs This depth of 1 seconds elapsed time is a the between direction simple system special- investigated time cost colmore in taxonomies. has only inference best Taxo- on incorporating paper learning than make and specificity machine is taxonomies. for reasoning between certainty .05)) chaining To rules This knowledge of information effective, ized inference time .05 second --- Command more of rules. [O.OO 0.451 time = ENTER world form the lections trade-off --- Much in and and the certainty. specificity and need yet to be explored. in which research the results hold much of promise. ACKNOWLEDGEMENTS <RESULT > [0.72 0.921 I would - time = ENTER .03 second --- Command or make > (? ((can-stop-in-time) using censor 0.017470836 query of system .03)) chaining depth of 0 seconds elapsed time <RESULT> --- for his guidance, [0.72 0.921 Command Cannot perform Minimum of this Kelly for help on an earlier D., Prade, inference query of system August, in requested time Ginsberg, time. is 0.018423745 set CONCLUSIONS It has of censored probabilistic the constraints. cations made tion. 242 / SCIENCE been semantics time that allows the when a system to conform a system in situations Examples John Wiegand on a and Jim implementation. H. “Combination and Propo- 1985, pp. 111-113. M.L. Dempster’s Austin, Texas, R.S., has numerous to time appli- decisions must with uncertain informa- and from medical (accepted 1984, pp. 126-129. August, Winston,P.H., “Variable 857, Artificial Laboratory, MIT, August, for AI Journal, Preci1985 1986). to adjust where range a Reasoning Rule” In Proc. AAAI-84. MIT AI Memo Intelligence formalgiven “Non-monotonic Using Michalski, rules of inferences Such in real shown production certainty Michalski .Ol)) sion Logic” ism and Prof. for comments gation of Uncertainty with Belief Fuctions” In Proc. IJCAI-85. Los Angeles, California, or make guaranteed V draft Dubois, > (? ((can-stop-in-time) paper, Wolf REFERENCES time limit too low --- ENTER like to thank Lisa be expert Shafer, G. A Mathematical Princeton Jersey, University 1976. Theory of Evidence. Press, Princeton, New