Angluin, Dana, Martins Krikis, Robert Sloan, and Gyorgy Turan. 1997. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211-- 255.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....algorithm that, given F as input and an membership oracle for an element f of F , is guaranteed to learn f using opt(F ) queries. Next, we look at the more general case. To study this case, we use a variant of the membership query model similar to that proposed by Angluin, Krikis, Sloan and Tur an [2]. Here, the range of the target f (our model of the user) and the functions in F (the experts) is an arbitrary nite set Y . As mentioned above, the algorithm is given an integer valued metric # on Y Y , and the distance between functions f and g is measured by # x#X #(f(x) g(x) The target ....

D. Angluin, M. Krikis, R. H. Sloan, and G. Tur an. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

....to selecting from among the possible pieces of generalization knowledge that might be relevant to the current example. Two important issues to be dealt with by a PBD system are the possibility that the user will want to respond I don t know or may make classification errors (cf. Angluin et al. [2]) In our case, errors could be introduced by outright incorrect annotations or by the user reconceptualizing the domain. Namely, over the course of defining several examples, the user may change what they believe to be the correct answer. Reconceptualizing the domain can lead to the changes in ....

D. Angluin, M. Krikis, R. Sloan, and G. Turan. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

....algorithm that, given F as input and an membership oracle for an element f of F , is guaranteed to learn f using opt(F ) queries. Next, we look at the more general case. To study this case, we use a variant of the membership query model similar to that proposed by Angluin, Krikis, Sloan and Tur an [2]. Here, the range of the target f (our model of the user) and the functions in F (the experts) is an arbitrary finite set Y . As mentioned above, the algorithm is given an integer valued metric on Y Y , and the distance between functions f and g is measured by P x2X (f(x) g(x) The ....

D. Angluin, M. Krikis, R. H. Sloan, and G. Tur an. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

....and equivalence queries (i.e. the exact learning model) 3] attracted a lot of attention. In particular, various concept classes were shown to be learnable in this model (e.g. 1, 2, 5, 12, 13, 26, 28, 14, 10, 9] and many others) In some of the above, and in several related papers (e.g. [3, 18, 6, 13, 8, 11, 14, 7]) the following common approach is used: the input space, f0; 1g n , is viewed as a lattice with the natural partial order, i.e. for u; v 2 f0; 1g n if u[i] v[i] for all i then u v (where u[i] is the i th bit of u) Then, the learning algorithm collects information about the target ....

D. Angluin, M. Krikis, R. E. Sloan, and G. Tur an. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

....queries) In a very early stage of the software design process, I don t know answers to membership queries might also be useful. However, we do not allow them here since their proper treatment would make the synthesis algorithm considerably more complicated. The reader is referred to [3] and the references given there for theoretical results concerning so called limited membership queries to which I don t know answers are possible. MAS contains an observation table T containing the current information about members and non members of the unknown language. The rows of T are ....

....her mind or accidentally gives an incorrect answer to a query. However, MAS always operates as if all the information so far obtained were correct. Hence, we do not apply such concepts as malicious omissions and errors or the corresponding inference algorithms introducted by Angluin et al. in [3]. This means that we have to update the observation table when incorrect information is detected in order to keep the observation table and auxiliary data structures consistent. This task is the subject of this chapter. Trie is a data structure for maintaining a set of strings over a given ....

D. Angluin, M. Krikis, R.H. Sloan, and G. Turan, Malicious omissions and errors in answers to membership queries. Mach. Learn. 28 (1997), 211-255.

....attention. In particular, various concept classes were shown to be learnable in this model (e.g. Ang87a, Ang87b, AFP92, BR92, Bsh93, RS93, SS93, Bsh95, BCV96, BBB 96] and many others) In some of the above, and in several related papers (e.g. Ang88, GM92, AHK93, Bsh93, AS94, BCGS95, Bsh95, AKST97] the following common approach is used: the input space, f0; 1g n , is viewed as a lattice with the An early version of this paper appeared in the proceedings of the 11th Workshop on Computational Learning Theory (COLT) pp. 294 302, July 1998. y Division of Engineering Applied ....

D. Angluin, M. Krikis, R. E. Sloan, and G. Tur'an. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

.... at the Technion, and by the German Israeli Foundation for scientific research and development (GIF) Ang87b, AFP92, RS93, Bsh93, BR95, SS96, BCV96, BBB 96, Bsh97] and many others) In some of the above, and in several related papers (e.g. Ang88, GM92, AHK93, Bsh93, AS94, BCGS95, Bsh97, AKST97] the following common approach is used: the input space, f0; 1g n , is viewed as a lattice with the natural partial order (i.e. for u; v 2 f0; 1g n if u[i] v[i] for all i then u v) Then, the learning algorithm collects information about the target function f by repeatedly looking for ....

D. Angluin, M. Krikis, R. E. Sloan, and G. Turan. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211--255, 1997.

....grant CCR 9208170, OTKA grant T 14228, and Phare TDQM grant 9305 02 1022 (ILP2 HUN) Email: U11557 uicvm.uic.edu. picks the subset of queries that receive a don t know response, and Angluin and Krik is studied a model where an adversary picks a subset of queries that receive the wrong response [2, 3, 22]. A practical motivation for studying noisy membership queries comes from the experiments of Lang and Baum [15] which confirm the intuition that erroneous answers to membership queries are more frequent close to the boundary of the target concept. Blum et al. 7] formulated interesting models ....

....the learner is allowed only a number of queries bounded by a polynomial in n, regardless of the frequency of responses. In general, there will be exponentially many instances in the boundary region even for constant r. Let us briefly compare this model to the limited membership query model [3, 22]. On the one hand, in the limited membership query model, the oracle is allowed to respond on any instance it chooses, not only on boundary instances. On the other hand, in that model the learner is allowed a number of queries polynomial in both n and the number of responses received. In ....

[Article contains additional citation context not shown here]

D. Angluin, M. Krik¸is, R. H. Sloan, and G. Tur'an. Malicious omissions and errors in answers to membership queries. Machine Learning. To appear.

No context found.

Angluin, Dana, Martins Krikis, Robert Sloan, and Gyorgy Turan. 1997. Malicious omissions and errors in answers to membership queries. Machine Learning, 28:211-- 255.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC