In this lecture we give several examples and lessons ues (see [2])? (Second) In our model we have a conlearned from research, development and experiments tinuous semantic, on the other side in [6] GAP was in the area of theory and applications of information not continuous (and general connection between these technology. We will try to describe a possible synergy two models was also an open problem). This was all of theory and application too. Namely, to describe solved in [2], introducing a model based on left conwhere practical needs bring new problems for theory tinuous conjunctors, with a weak form of border conand where theory helps to formulate methods, which dition (without associativity and commutativity) and should be verified in practice. with body aggregation . We have shown that FLP is In the theoretical part we will mention research on (in a sense) equivalent to GAP. correctness and completeness of fuzzy logic program- In the application part we make difference between ming [1,2] and various measures for evaluating success. case studies and use cases. Case studies include deIn applications we mention acquaintance with devel- scriptions of systems that have been deployed within opment and experiments of preferential querying and an organization, and are now being used within a prouser dependent top-k answers [3]. In all of these it also duction environment. Use cases include examples depends on whether our task is deductive (querying), where an organization has built a prototype system, inductive (learning) or abductive. In practice, it is im- but it is not currently being used by business funcportant to have a user behavior model (for many dif- tions [18]. Repeatability of experiments is also an isferent users). To see what is (can be, must be) done au- sue, see e.g. [19]. So our applications here are not true tomatically (trained, assisted, unsupervised, . . . ) and deployed applications, they are rather experiments what by human, what is domain dependent and what (use case prototypes) and repeatability is not always is generic. Our fuzzy model is not a mere generaliza- fully satisfactory enabled. tion from two values to many values. The key point of Main impulse for these considerations came from our study here is our understanding of fuzzy value as a referee refuting our paper in an application oriented preference degree. Using fuzzy as preferences enlight- conference. He/She asked - where from do you have ens phenomena which in a two valued world are not rules of your FLP? So far main motivation in fuzzy visible at all. were toy examples with tall Swedes and young basFrom mathematical point of view one can general- ketball players. Our motivation was real life examples ize LP to many valued logic. Also here we face several where fuzzy degree is the degree of user preference challenges. Should our rules be implications or clauses, (and no more fuzzy linguistic variables with modifiers should our computation be refutation or query answer- like very tall . . . ). A well developed counterpart of this ing, is unification touched by this or not? In two valued is already in preferential querying, where only top-k logic these are equivalents, in [1] we have developed a most preferred answers are interesting. Main contrimodel of ([0,1] valued) fuzzy logic programming FLP bution here was made by R. Fagin (see e.g. [10]), who with implicative rules and computation based on back- (in a datalog setting, without function symbols) asward usage of modus pones (and possible extension sumed we have objects in several lists repeatedly orwith fuzzy similarity in [4]). dered by different attributes (local) preferences and Concerning implementation of this system, gave an original optimal algorithm for top-k for this M. Lieskovsky has constructed in [5] a fuzzy War- setting. This direction was further investigated ren abstract machine. Our system enables new form by P. Gursky´ who has implemented and experimented of cuts for threshold queries (see [1]). with several heuristics (see e.g. [7]). V. Vanekov´a has Further development went in two directions. developed several knowledge representation models for (First) What happens in finitely valued case when dif- this ([8]). ferent attributes take different number of truth val- But referee asked where from do you have those rules (Fagin assumes we have the (query) rules, with ⋆ This work was partially supported by Czech Grant local preferences and global aggregation (in good conAgency GA CR under number 202/10/0761. cordance with our result FLP = GAP)? So now the
question is, where from we have local preferences (user 8. V. Vanekova, P. Vojtas: Comparison of scoring and preferences on attributes represented by a fuzzy (rank- order approach in description logic EL(D). SOFSEM ing) set) and where from do we have combination (ag- 2010: 709{720. gregation] function giving the global preference? This 9. A. Eckhardt, T. Horvath, P. Vojtas: Learning different is an inductive task. With T. Horv´ath, A. Eckhardt we user profile annotated rules for fuzzy preference top-k have developed several inductive models (see e.g. [9]). querying. SUM 2007, 116{130. 10. R. Fagin, A. Lotem, M. Naor: Optimal aggregation al
Moreover a practical problem occurred. Well, as- gorithms for middleware. J. Comput. Syst. Sci. 66 (
For theoretical part it is now clear that equality tomating process connecting web and user. URSW of fuzzy sets (correct answers and computed answers) (LNCS Vol.) 2008, 207{223. is not a good measure and correctness and complete- 16. J. Dedek, P. Vojtas: Fuzzy classification of web reness results have to be reconsidered with some order ports with linguistic text mining. Web Intelligence/IAT violation/concordance measures. Workshops 2009, 167{170.
Further, from an experimental point of view, we 17. R. Novotny, P. Vojtas, D. Maruscak: Information
exwent in direction of web information extrac- traction from web pages. Web Intelligence/IAT
Worktion ([16,17]), because it is also interesting to know 18. Ssheompasn2t0ic09,W12e1b{12C4a.se Studies and Use Cases,
where are all these data from (after where are rules http://www.w3.org/2001/sw/sweo/public/UseCases/.
from). Situation connecting web and user is heavily 19. S. Manegold et al.: Repeatability &
workinfluences by uncertainty, starting research is done ability evaluation of SIGMOD 2009.
SIGin [15]. MOD Record 38 (
We can conclude, that synergy between theoretical http://www.sigmod08.org/sigmod call papers.shtml and applied (experimental) research and development sub 6. was beneficial for both of them.