Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Sydney's_Scientific_Beginnings:_William_Dawes'_Obs.txt	While Dawes was clearly a bright and intelligent young man, Maskelyne’s choice of him as his emissary nevertheless seems odd, given his working-class background. As the nation’s pre-eminent astronomer, Maskelyne was not short of well-qualified, university- educated volunteers, so Dawes must indeed have come with the highest praise. Dawes’ successful appointment by the Home Secretary, Thomas Townshend, First Viscount Sydney, and the Board of Longitude is even more surprising given the enormous importance afforded to establishing the new colony. Other luminaries, including Banks and Sir Evan Nepean (1752–1822), Permanent Undersecretary of State for the Home Department, also vied for patronage. Indeed, Maskelyne seems to have outwitted his nemesis Banks in securing Dawes’ loyalty (e.g., Gibson, 2012).
Processed_Sydney's_Scientific_Beginnings:_William_Dawes'_Obs.txt	From our current perspective, it may seem odd that sending an astronomer to the yet-to-be-established colony—even if he was also a Marine, a competent engineer, surveyor and explorer—was considered a high priority. Shortages of labour, materials and food would hamper the colony’s early development and keep Governor Phillip occupied, yet Dawes was meanwhile commissioned to establish an astronomical observatory.
Processed_Sydney's_Scientific_Beginnings:_William_Dawes'_Obs.txt	The focus of this article is on Dawes’ observatory in New South Wales, 5 the developments leading up to its foundation (Section 2), the instruments it was equipped with and the role it played in the social fabric of the new colony (Section 3). While some of this story is reasonably well known, our aim here is to provide a comprehensive account of the history and operations of the first permanent6 British astronomical observatory in the southern hemisphere. We have uncovered a number of new insights, and we have also established the observatory’s site beyond any reasonable doubt based on a careful assessment of the historical record (Section 4). Our historical research has been based on a wide range of records. Among the online documents we consulted were the full set of papers of the Board of Longitude and eyewitness accounts of the First Fleet’s voyage. In addition, we explored the online and physical holdings of the State Library and State Archives of New South Wales, the library of the Royal Australian Historical Society (RAHS), all in Sydney, and the National Library of Australia in Canberra.
Processed_Sydney's_Scientific_Beginnings:_William_Dawes'_Obs.txt	In the late eighteenth century, the dominant sea-faring European nations continued to open up commercial shipping routes to ever more distant destinations, yet a reliable means of geographic position determination at sea remained elusive. Determination of one’s latitude was relatively straightforward: one simply needed to measure the height of the Sun or those of one or more bright stars at their meridian passage, corrected for seasonal variations, and a latitude measurement would follow naturally.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	The spin-isospin response of stable and exotic nuclei is investigated in the framework of the proton-neutron relativistic quasiparticle time-blocking approximation (pn-RQTBA). Based on the Covariant Density Functional Theory, this method extends the proton-neutron Relativistic Quasi- particle Random-Phase Approximation (pn-RQRPA) by including the coupling between single quasiparticles and collective nuclear vibrations. In the charge-exchange channel, this coupling generates a time-dependent effective interaction between proton and neutron quasiparticles. The particle-hole component of this interaction adds to the static pion and rho-meson exchange, while the particle-particle component provides a microscopic and consistent proton-neutron pairing in- teraction. We ﬁnd that such dynamical effects induce fragmentation and spreading of the Gamow- Teller transition strength which are important for a better agreement with the experimental mea- surements and for an accurate description of β -decay rates. The new developments include the coupling of single nucleons to isospin-ﬂip vibrations in doubly-magic nuclei. We ﬁnd that these phonons can have a non-negligible effect on β -decay half-lives and "quenching" of the strength.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	c(cid:13) Copyright owned by the author(s) under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Nuclear transitions involving transfer of isospin have many applications in nuclear and parti- cle physics, as well as in astrophysics as they determine weak-interaction rates governing r-process nucleosynthesis and stellar evolution. Such types of modes as Fermi, Gamow-Teller (GT) or spin- dipole (SD) transitions are very sensitive to the spin-isospin dependence of the nucleon-nucleon interaction and have been extensively studied in many theoretical methods (see e.g. Refs [1] to [20]). In the Relativistic Random-Phase Approximation (RRPA) based on the Relativistic Mean- Field (RMF), where the exchange interaction is not explicitly treated, isospin-transfer modes are fully determined by isovector rho- and pion-exchange. While the proton-neutron (R)RPA (pn- (R)RPA) usually describes the position of GT and SD giant resonances quite well, extensions are necessary to include spreading mechanism essential for the description of width and details of the transition strength. In open-shell nuclei with superﬂuid pairing correlations, it has been shown that an isoscalar residual pairing interaction is important for an accurate description of the GT strength and weak-interaction rates [5, 6, 7, 11]. In proton-neutron (Relativistic) Quasiparticle-RPA (pn- (R)QRPA) studies, the strength of this interaction is however usually ﬁtted to β -decay half-lives which lowers the predictive power of the method. In this work we study the GT response of a few mid-mass nuclei within the Relativistic Quasiparticle- Vibration Coupling (RQVC) framework presented in Ref. [20]. This approach, applied in the proton-neutron Quasiparticle Time-Blocking Approximation (pn-RQTBA), extends the pn-RQRPA by accounting for the interplay between quasiparticle and collective degrees of freedom, and natu- rally generates a dynamical pn pairing interaction, without extra empirical factor. We compare our results to the pn-RQRPA and to the available experimental data. For the ﬁrst time we include the coupling of nucleons to isospin-ﬂip vibrations in doubly-magic nuclei. We investigate their effect on the quenching of the GT strength and β -decay half-lives.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	The sum (cid:101)Vρ + (cid:101)Vπ is the ﬁnite range isovector meson-exchange interaction, while (cid:101)Vδπ denotes the zero-range Landau-Migdal term taken with parameter g(cid:48) = 0.6, as the exchange interaction is not treated in the present work [12]. When considering this static interaction alone, one gets back to the pn-RQRPA. The QVC amplitude Φ(ω) is represented in Fig. 1 in terms of Feynman diagrams. This dynamical interaction introduces 1(q)p-1(q)h⊗phonon conﬁgurations and is responsible for damping of the transition strength. A detailed expression of W (ω) can be found in Ref. [20]. In the following the extension of the pn-RQRPA including QVC effects in the TBA is referred to as pn-RQTBA.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	For such unnatural parity modes the pion gives almost the full contribution of the static meson exchange interaction (2.5). As it is absent in the RHB ground-state, the pion is considered with the free-space coupling constant f 2 4π = 0.08, which makes GT a great study-case where no double- counting effect is expected when going beyond the pn-RQRPA description. In this work we calcu- late the response of stable and neutron-rich nuclei to the ˆF GT − operator, using the numerical scheme described in Ref. [20]. Starting from the RHB approximation with the NL3 meson parametrization [23] and a monopole-monopole pairing force [22], the set of phonons that are coupled to the quasi- particles is calculated in the RQRPA. As a ﬁrst step, we consider the coupling to non-isospin-ﬂip phonons with Jπ = 2+, 3−, 4+, 5−, 6+ in an energy window of 30 MeV. The GT response is then calculated by solving Eq. (2.4), and the strength distribution is obtained from Eq. (2.1). In Ref. [20] we calculated the GT− response of neutron-rich Nickel isotopes within this approach. In Figs. 2a and 2b we show the strength distributions in 68Ni and 78Ni calculated with a smear- ing parameter ∆ = 200 keV. For comparison we show the results obtained at the pn-RQRPA level in blue, while the full calculations including QVC (pn-RQTBA) are displayed with red lines. In order to obtain the excitation spectrum with respect to the daughter ground-state, all distributions have been shifted by the binding energy difference ∆B = B(parent) − B(daughter) calculated in the RHB approximation. Clearly, the QVC induces fragmentation and spreading of the pn-RQRPA transition strength over a larger energy range including the low-energy region. We saw that such effects resulted in a considerable improvement of the β -decay half-lives as compared to the pn- RQRPA description [20]. We also show in Fig. 2a the low-lying strength distribution in 68Ni cal- culated with ∆ = 20 keV in order to disentangle the low-lying states. Experimentally it is known that the ground-state of 68Cu has angular momentum and parity 1+ [24]. The shift induced by QVC clearly improves the position of the 1+ 1 state, as it is found to lie at 330 keV versus 3.88 MeV in pn-RQRPA. We remind that we do not introduce any static proton-neutron pairing. In pn-(R)QRPA studies such pairing is usually introduced through an attractive residual isoscalar pn interaction in the particle-particle channel which is ﬁtted to e.g. β -decay half-lives [5, 6, 7, 11]. When the QVC interaction of Fig. 1 is taken into account, pn pairing appears naturally due to the presence of particle-like pairing which generates the particle-particle component of the dynamical pn interaction Φ(ω). This can explain why the QVC induces a similar effect on the low-energy strength as the phenomenological static pn pairing.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Figure 2: GT− strength distribution in a few stable and neutron-rich nuclei. The pn-RQRPA results are shown with blue lines while the pn-RQTBA results (with QVC) are shown with red lines.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	We show in Figs. 2c and 2d the GT− strength distributions in 90Zr and 48Ca which have been measured experimentally. The strength in 90Zr has been calculated using a smearing ∆ = 1 MeV, to simulate the experimental resolution of Ref. [25]. When QVC is included, a nice agreement with the data is obtained up to an excitation energy of ∼ 25 MeV. At higher energy the experimental strength also contains the isovector spin-monopole response [25]. The GT− strength in 48Ca has been measured with a higher energy resolution corresponding to ∆ = 210 keV in Ref. [26]. At such high resolution, the pn-RQRPA is not able to reproduce the details of the experimental spec- trum. The position of the ﬁrst low-lying state and giant resonance region are also overestimated by ∼ 3 − 4 MeV. The fragmentation induced by QVC leads to the appearance of ﬁner details in the transition strength, which is also shifted down in accordance with the data. The peak around ∼ 13 MeV however still lacks fragmentation.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Up to now the phonon spectrum coupled to the quasiparticles was limited to non-isospin-ﬂip phonons with natural parity. Ref. [27] however showed that the presence of low-lying isospin-ﬂip modes can have an impact on the single-nucleon shell structure. Recent developments now make it possible to include the coupling to charge-exchange phonons in the description of the response in doubly magic nuclei. Such vibrations introduce new diagrams in the QVC interaction that are dis- played in Fig. 3. We note that due to charge conservation they only generate additional self-energy insertions, and do not appear in the form of a phonon-exchange. Moreover, we see that these new self-energy terms involve proton-neutron particle-particle elements of the particle-vibration cou- pling vertex, and thus can also be interpreted as a (virtual) energy-dependent pn pairing interaction in doubly-magic nuclei.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Figure 3: QVC interaction in the proton-neutron channel. The wavy lines represent isoscalar (non- isospin-ﬂip) phonons while the springs represent isovector (isospin-ﬂip) phonons.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Figure 4: GT− strength distribution in 78Ni and corresponding cumulative sum of B(GT−). The results are shown at different theoretical levels: at the pn-RQRPA level (in blue), at the pn-RQTBA level with coupling to isoscalar (IS) phonons in a QVC energy window Ewin = 30 MeV in dashed red, at the pn-RQTBA level with coupling to IS and charge-exchange phonons in a QVC energy window Ewin = 30 MeV in plain red, and at the pn-RQTBA level with coupling to IS and charge- exchange phonons in a QVC energy window Ewin = 100 MeV in black.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	phonons brings further fragmentation of the peak at ∼ 13 MeV which is counterbalanced by an increase of the strength around ∼ 7 MeV. We now see a clear state at ∼ 18.3 MeV which is ∼ 1.3 MeV too high than the one seen experimentally. The ﬁrst state around 2 MeV is shifted down when introducing charge-exchange phonons and increasing the QVC window Ewin to a value of 100 MeV. Around 30 MeV, which is the excitation energy reached experimentally, we note again from Fig. 6b that the pn-RQRPA distribution has almost saturated as it reaches ∼ 98% of the total GT− strength in the particle sector. The strength with QVC (in black) introduces some quenching since it exhausts only ∼ 91% of that value. Such redistribution of the transition strength is however not sufﬁcient to reach the experimental value of only 71%.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Figure 5: Half-life of 78Ni in seconds. The experimental value is taken form Ref. [28]. The error bars are experimental ones.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	Figure 6: GT− strength in 48Ca in the particle sector and cumulative sum of the B(GT−) values.
Processed_Relativistic_approach_to_nuclear_spin-isospin_exci.txt	In this work we applied the pn-RQTBA approach to the description of Gamow-Teller (GT) transitions in a few mid-mass nuclei. We found that the coupling between quasiparticles and col- lectives vibrations, which introduces complex 1qp-1qh⊗phonons conﬁgurations, systematically induces fragmentation and spreading of the pn-RQRPA transition strength, and leads to a more de- tailed description of the GT states, in better accordance with the data. We included for the ﬁrst time the coupling to charge-exchange vibrations in doubly magic nuclei and found that such phonons can have a non-negligible impact on the β -decay half-lives and on the quenching of the GT strength. When the experimental resolution is high, it is however clear that the RQVC framework imple- mented in the present Time-Blocking Approximation is often not sufﬁcient to reproduce the very ﬁne details of the strength. In order to further improve the description of spin-isospin modes we plan for the future to include higher-order conﬁgurations in the response [30] and in the ground state.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In inductive learning of a broad concept, an algorithm should be able to distinguish concept examples from exceptions and noisy data. An approach through recursively ﬁnding patterns in exceptions turns out to correspond to the problem of learning default theories. Default logic is what humans employ in common- sense reasoning. Therefore, learned default theories are better understood by humans. In this paper, we present new algorithms to learn default theories in the form of non-monotonic logic programs. Experiments reported in this paper show that our algorithms are a signiﬁcant improvement over traditional approaches based on inductive logic programming. Under consideration for acceptance in TPLP.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Predictive models produced by classical machine learning methods are not comprehensible for humans because they are algebraic solutions to optimization problems such as risk minimization or data likelihood maximization. These methods do not produce any intuitive description of the learned model. This makes it hard for users to understand and verify the underlying rules that govern the model. As a result, these methods do not produce any justiﬁcation when they are applied to a new data sample. Also, extending the prior knowledge1 in these methods requires the entire model to be re-learned. Additionally, no distinction is made between exceptions and noisy data. Inductive Logic Programming (Muggleton 1991), however, is one technique where the learned model is in the form of logic programming rules (Horn clauses) that are more com- prehensible and that allows the background knowledge to be incrementally extended without requiring the entire model to be relearned. This comprehensibility of symbolic rules makes it easier for users to understand and verify the resulting model and even edit the learned knowl- edge.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Given the background knowledge and a set of positive and negative examples, ILP learns the- ories in the form of Horn logic programs. However, due to the lack of negation-as-failure, Horn clauses are not sufﬁciently expressive for representation and reasoning when the background knowledge is incomplete.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Additionally, ILP is not able to handle exception to general rules: it learns rules under the as- sumption that there are no exceptions to them. This results in exceptions and noise being treated in the same manner. Often, the exceptions to the rules themselves follow a pattern, and these exceptions can be learned as well. The resulting theory that is learned is a default theory, and in most cases this theory describes the underlying model more accurately. It should be noted that default theories closely model common sense reasoning as well (Baral 2003). Thus, a de- fault theory, if it can be learned, will be more intuitive and comprehensible for humans. Default reasoning also allows us to reason in absence of information. A system that can learn default theories can therefore learn rules that can draw conclusions based on lack of evidence, just like humans. Other reasons that underscore the importance of inductive learning of default theories can be found in Sakama (Sakama 2005) who also surveys other attempts in this direction.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	As an example, suppose we want to learn the concept of ﬂying ability of birds. We would like to learn the default rule that birds normally ﬂy, as well as rules that capture exceptions, namely, penguins and ostriches are birds that do not ﬂy. Current ILP systems will be thrown off by the exceptions and will not discover any general rule: they will just either enumerate all the birds that ﬂy or cover the positive examples without caring much about the falsely covered negative examples. Other algorithms, such as FOIL, will induce rules that are non-constructive and thus not helpful or intuitive.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In this paper, we present two algorithms for learning default theories (i.e., non-monotonic logic programs), called FOLD (First Order Learner of Default) and FOLD-R, to handle categor- ical and numeric features respectively. Unlike traditional ILP systems that learn standard logic programs (i.e., no negation is allowed), our algorithms learn non-monotonic stratiﬁed logic pro- grams (that allow negation-as-failure). Our algorithms are an extension of the FOIL algorithm (Quinlan 1990) and support both categorical and numeric features. Also, the FOLD and FOLD-R learning algorithms can learn recursive rules. Whenever needed, our algorithms introduce new predicates. The language bias (Mitchell 1980) also contains arithmetic constraints of the form {A ≤ h, A ≥ h}. The algorithms have been implemented and tried on variety of datasets from the UCI repository. They have shown excellent results that are presented here as well.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	The default theories that we learn using our algorithm, as well as the background knowledge used, is assumed to follow the stable model semantics2. Stable model semantics, and its real- ization in answer set programming(ASP), provides an elegant mechanism for handling negation in logic programming (Gelfond and Lifschitz 1988). We assume that the reader is familiar with ASP and stable model semantics (Baral 2003).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	This paper makes the following contributions: We propose a novel concrete algorithm to learn default theories automatically in the absence of complete information. The proposed algorithm, unlike the existing ones, is able to handle the numeric features without discretizing them ﬁrst, and is also capable of handling non-monotonic background knowledge. We provide both quali- tative and quantitative results from standard UCI datasets to support the claim that our algorithm discovers more accurate as well as more intuitive rules compared to the conventional ILP sys- tems.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	the experiments and results. Section 7 discusses related research. Section 8 discusses our future research direction and ﬁnally in Section 9 we conclude.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	a theory T for which covers(T , E +, B) = E + and covers(T , E −, B) = /0.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Our algorithm to learn default theories is an extension of the FOIL algorithm (Quinlan 1990). FOIL is a top-down ILP system which follows a sequential covering approach to induce a hy- pothesis. The FOIL algorithm is summarized in Algorithm 1. This algorithm repeatedly searches for clauses that score best with respect to a subset of positive and negative examples, a current hypothesis and a heuristic called information gain (IG).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	candidate literals for any existing candidate L. This approach leads to learning predicates that do not capture the concept accurately as shown in the following example.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Example 3.1 B, E + are background knowledge and positive examples respectively with CWA and the concept to be learned is ﬂy.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	either noisy data samples or exceptions to the so far learned rule. As (Srinivasan et al. 1996) dis- cuss, there is no pattern distinguishable in noise, whereas, in exceptions, there may exist a pattern that can be described using the same language bias. This can be viewed as a subproblem to (re- cursively) ﬁnd the rules governing a set of negative examples. To achieve that aim, FOLD swaps the current positive and negative examples and recursively calls the FOLD algorithm to learn the exception rule(s). Each time a rule is discovered for exceptions, a new predicate ab(V1, ...,Vn) is introduced. To avoid name collision, FOLD appends a unique number at the end of the string ab to guarantee the uniqueness of the invented predicates.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In case of noisy data or in the presence of uncertainty due to the lack of information, it turns out that there is no pattern to learn. In such cases, FOLD enumerates the positive examples for two purposes: ﬁrst, this is essential for the training algorithm to converge, second, it helps to detect noisy data samples.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Algorithm 2 shows a high level implementation of the FOLD algorithm. In lines 1-8, function FOLD, serves as the FOIL outer loop. In line 3, FOLD starts with the most general clause (e.g. f ly(X) ← true.). In line 4, this clause is reﬁned by calling the function SPECIALIZE. In lines 5-6, set of positive examples and set of discovered clauses are updated to reﬂect the newly dis- covered clause. In lines 9-29, the function SPECIALIZE is shown. It serves as the FOIL inner loop. In line 12, by calling the function ADD BEST LITERAL the “best” positive literal is cho- sen and the best IG as well as the corresponding clause is returned. In lines 13-24, depending on the IG value, either the positive literal is accepted or the EXCEPTION function is called. If, at the very ﬁrst iteration, IG becomes zero, then a clause that just enumerates the positive examples is produced. A ﬂag called just started is used to differentiate the ﬁrst iteration. In lines 26-27, the sets of positive and negative examples are updated to reﬂect the changes of the current clause. In line 19, the EXCEPTION function is called while swapping the E +, E −.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In line 31, we ﬁnd the “best” positive literal that covers more positive examples and fewer negative examples. Again, note the current positive examples are really the negative examples and in the EXCEPTION function, we try to ﬁnd the rule(s) governing the exception. In line 33, FOLD is recursively called to extract this rule(s). In line 34, a new ab predicate is introduced and in lines 35-36 it is associated with the body of the rule(s) found by the recurring FOLD function call in line 33. Finally, in line 38, default and exception are attached together to form a single clause.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	abnormalities namely, AB. In line 38, the negated exception (i.e not ab0(X)) and the default rule’s body (i.e bird(X)) are compiled together to form the clause f ly(X) ← bird(X), not ab0(X). Note, in two different cases enumerate is called. First, at very ﬁrst iteration of specialization if IG is zero for all the positive literals. Second, when the Exception routine fails to ﬁnd a rule gov- erning the negative examples. Whichever is the case, corresponding samples are considered as noise. The following example shows a learned logic program in presence of noise. In particular, it shows how enumerate function in FOLD works: It generates clauses in which the variables of the goal predicate can be uniﬁed with each member of a list of the examples for which no pattern exists.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	f ly(X) ← bird(X), not ab0(X). f ly(X) ← member(X, [ jet]). ab0(X) ← penguin(X).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	FOLD recognizes jet as a noisy data. member/2 is a built-in predicate in SWI-Prolog to test the membership of an atom in a list.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Example 4.2 Birds and planes normally ﬂy, except penguins and damaged planes that can’t. There are super penguins who can, exceptionally, ﬂy.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	bird(X) ← penguin(X). penguin(X) ← superpenguin(X). bird(a). bird(b). penguin(c). penguin(d). superpenguin(e). superpenguin( f ). cat(c1). plane(g). plane(h). plane(k). plane(m). damaged(k). damaged(m). f ly(a). f ly( f ).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	f ly(X) ← plane(X), not ab0(X). f ly(X) ← bird(X), not ab1(X). f ly(X) ← superpenguin(X). ab0(X) ← damaged(X). ab1(X) ← penguin(X).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Proof It sufﬁces to show that the size of E + on every iteration of the FOLD function decreases (at line 5) and since E + is a ﬁnite set, it will eventually becomes empty and the while loop terminates. Equivalently, we can show that every time the SPECIALIZE function is called, it terminates and a clause ˆc that covers a non-empty subset of E + is returned. Inside the SPECIALIZE function, if E − is empty, then the function returns its input clause and the theorem trivially holds. Otherwise, two cases might happen: First, SPECIALIZE produces a clause which enumerates E + and covers no negative example. In such a case it returns immediately and the theorem trivially holds. Sec- ond, SPECIALIZE calls the EXCEPTION function which may lead to a chain of recursive calls on FOLD function. In this case it sufﬁces to show that on a chain of recursive calls on FOLD, the size of function argument i.e. E + decreases each time a new call to FOLD occurs. That’s indeed the case because every time a literal is added to the current clause at line 12, it covers fewer neg- ative examples from E − which in turn becomes the new E + as the EXCEPTION function and subsequently the FOLD function is recursively called. Therefore, on consecutive calls to FOLD function, the size of input argument E + is decreased until it eventually terminates.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Theorem 2 The FOLD algorithm always learns a hypothesis that covers no negative example (soundness).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Proof It follows from the Theorem 1 that every loop in the algorithm (and, subsequently the algorithm) terminates. In particular, the while loop inside the function SPECIALIZE terminates as soon as the negated loop condition (i.e., the number of negative examples covered by the rule being dis- covered equals zero) starts to hold. Since, for every learned rule, no negative example is covered, it follows that the FOLD algorithm learns a hypothesis which covers no negative example.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Theorem 3 The FOLD algorithm always learns a hypothesis that covers all positive examples (complete- ness).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	ILP systems have limited application to data sets containing a mix of categorical and numerical features. A common way to deal with numerical features is to discretize the data to qualitative values. This approach leads to accuracy loss and requires domain expertise. Instead, we adapt the approach taken in the well-known C4.5 algorithm (Quinlan 1993). This algorithm is ranked no. 1 in the survey paper “Top 10 algorithms in datamining”, (Wu et al. 2007). For a numeric feature A, constraints such as {A ≤ h, A > h} have to be considered where the threshold h is found by sorting the values of A and choosing the split between successive values that maximizes the information gain. In our FOLD-R algorithm that we propose and describe next, we perform the same method for a set of operators {<, ≤} and pick the operator and threshold which maximizes the information gain. Also, we need to extend the ILP language bias to support the arithmetic constraints.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	for numerical features, we deﬁne a predicate that contains an extra variable which always pairs with a constraint. For example to extend the language bias for a numeric quantity “age” we could deﬁne predicates of the form age(a, b) in the background knowledge, and the candidate to specialize a clause might be as follows: age(X, N), N ≤ 5. However, the predicate age/2 never appears without the corresponding constraint.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Algorithm 3 illustrates the high level changes made to FOLD, in order to obtain the FOLD-R algorithm. The function test categorical, as before, chooses the best categorical literal to spe- cialize the current clause. The function test numeric chooses the best numeric literal as well as the best arithmetic constraint and threshold with the highest IG. In line 5, if neither one leads to a positive IG, EXCEPTION is tried. If exception also fails, then enumerate is called. Otherwise, IGs are compared and whichever is greater, the corresponding clause is chosen as the specialized clause of the current iteration.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Example 5.1 Table 1 adapted from (Quinlan 1993) is a dataset with numeric features “temperature” and “hu- midity”. “Outlook” and “Windy” are categorical features. Our FOLD-R algorithm, for the goal play(X), and positive examples shown as records with label “Play”, and negative examples shown as records with label “Don’t Play” outputs the following clauses: play(X) ← overcast(X). play(X) ← temperature(X, A), A ≤ 75, not ab0(X). ab0(X) ← windy(X), rainy(X). ab0(X) ← humidity(X, A), A ≥ 95, sunny(X).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	FOLD-R results suggest an abnormal day to play is either a rainy and windy day or a sunny day with humidity above 95%.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	good cont(X) ← wage inc f (X, A), A > 2, not ab0(X). good cont(X) ← holidays(X, A), A > 11. good cont(X) ← hplan hal f (X), pension(X). ab0(X) ← no long disability hel p(X). ab0(X) ← no pension(X).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	poisonous(X) ← ring type none(X). poisonous(X) ← spore print color green(X). poisonous(X) ← gill size narrow(X), not ab2(X). poisonous(X) ← odor f oul(X). ab0(X) ← population clustered(X). ab0(X) ← stalk sur f ace below ring scaly(X). ab1(X) ← stalk shape enlarging(X). ab2(X) ← gill spacing crowded(X), not ab1(X). ab2(X) ← odor none(X), not ab0(X).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Note, the induced theory has nested exceptions. This nesting happens as a result of ﬁnding pat- terns for negative examples, which makes the FOLD algorithm perform more recursive steps until no covered negative example is left. Table 2 compares the accuracy of FOLD-R algorithm against that of ALEPH (Srinivasan 2001). The examples have been picked from well-known standard datasets for some of which ALEPH exhibits low test accuracy. In most cases, FOLD-R accuracy outperforms ALEPH. The experi- ments suggest that when absence of a particular feature value plays a crucial role in classiﬁcation, our algorithm shows a meaningful higher accuracy. This follows from the fact that the classical ILP algorithms only make use of existent information as opposed to negation-as-failure in which a decision is made based on the absence of information. As an example, in the credit-j dataset, our algorithm generates 4 rules with abnormality predicates. These rules cover positive examples which without abnormality predicates would have remained uncovered. However, in Bridges and Ecoli where no abnormality predicate is introduced by our algorithm, both ALEPH and FOLD-R end-up with almost the same accuracy.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Even in cases where no improvement over accuracy is achieved, our default theory approach leads to simpler and more intuitive rules. As an example, in case of Mushroom, other ILP sys- tems, including ALEPH and FOIL, would produce 9 rules with 2 literals each in the body to cover all the positives, while our FOLD algorithm, produces 3 single-literal rules and 1 rule with 2 literals in which the second literal takes care of the exceptions.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Additionally, since ASP systems have to ground the predicates to produce the answer set, intro- ducing numeric data in background knowledge and also in the language bias is prohibited.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In a different line of research, (Dimopoulos and Kakas 1995), describes an algorithm to learn exception using the patterns in the negative examples. However, they don’t make any use of NAF as the core notion of reasoning in the absence of complete information. Instead, their algorithm learns a hierarchical logic program, including classical negation, in which the order of rules prioritize their application and, therefore, it is not naturally compatible with standard Prolog.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	The idea of swapping positive and negative examples to learn patterns from negative examples has ﬁrst been discussed in (Srinivasan et al. 1996) where a bottom-up ILP algorithm is proposed to specialize a clause after it has already been generalized and still covers negative examples. Similarly, (Inoue and Kudoh 1997) proposes a bottom-up algorithm in two phases: First, produc- ing monotonic rules via a standard ILP method, then specializing them by introducing negated literals to the body of the rule. In contrast, we believe our FOLD algorithm that takes a top-down approach learns programs that have a better ﬁt, thanks to its support for numeric features and its better scalability. Lack of both are inherent problems in bottom-up ILP methods.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	ALEPH (Srinivasan 2001) is one of the most widely used ILP systems that uses a bottom-up generalization search to induce theories that covers the maximum possible positive examples. However, since the induced theory might be overly generalized, there is an option to reﬁne the theory by introducing abnormality predicates that rule out negative examples by specializing an overly generalized rule. This specialization step is manual and unlike our algorithm, no automa- tion is offered by ALEPH. Also, ALEPH does not support numeric features.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	XHAIL (Ray 2009) is another non-monotonic ILP framework which integrates abductive, de- ductive and inductive forms of reasoning to form sets of ground clauses called kernel set and then generalize it to learn non-stratiﬁed logic programs.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	In a different line of research, (Corapi et al. 2012) approaches the non-monotonic ILP problem by incorporating the power of modern ASP solvers to search for an optimal hypothesis among the generated so-called skeleton rules and a set of abducibles associated to them. The last two approaches do not scale up as the language bias grows.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	One advantage of our FOLD-R algorithm over the existing systems is the ability to handle non- monotonic background knowledge. Conventional ILP systems permit only standard Prolog pro- grams to represent background knowledge. In contrast, our FOLD-R algorithm, once integrated with a top down answer set programming system like s(ASP) (Marple and Gupta 2012), will permit the background knowledge to be represented as an answer set program. The idea of lever- aging the added expressiveness of non-monotonic logic-based background knowledge has been discussed in (Seitzer 1997). However, because Seitzer’s method is based on grounding the back- ground knowledge and then computing its answer sets, it is not scalable. Using a query-driven system allows our learning algorithm to be scalable. Extending our algorithms to allow non- monotonic logic-based background knowledge with multiple stable models is part of the future work.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	Our eventual goal is to develop a uniﬁed framework for learning default theories: (i) in which we can learn hypotheses that are general answer set programs (i.e., these learned answer set programs may not be stratiﬁed), and (ii) that work with background knowledge that may be represented as a non-stratiﬁed answer set program. Also, the optimality of the learned hypothesis due to greedy nature of information gain heuristic is not guaranteed and changing the search strategy from A* to better algorithms such as iterative deepening search is subject of future research.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	tion, FOLD and FOLD-R are able to ﬁnd patterns of information that is absent and express it via a default theory, i.e., as a non-monotonic logic program.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	ALBERT, E., Ed. 2013. Logic-Based Program Synthesis and Transformation, 22nd International Sympo- sium, LOPSTR 2012, Leuven, Belgium, September 18-20, 2012, Revised Selected Papers. Lecture Notes in Computer Science, vol. 7844. Springer.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	CORAPI, D., RUSSO, A., AND LUPU, E. 2012. Inductive Logic Programming in Answer Set Programming.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	INOUE, K. AND KUDOH, Y. 1997. Learning extended logic programs. In Proceedings of the 15th Inter- national Joint Conference on Artiﬁcal Intelligence - Volume 1. IJCAI’97. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 176–181.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	LAVRAC, N. AND WROBEL, S., Eds. 1995. Machine Learning: ECML-95, 8th European Conference on Machine Learning, Heraclion, Crete, Greece, April 25-27, 1995, Proceedings. Lecture Notes in Com- puter Science, vol. 912. Springer.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	QUINLAN, J. R. 1990. Learning logical deﬁnitions from relations. Machine Learning 5, 239–266. QUINLAN, J. R. 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann. RAY, O. 2009. Nonmonotonic abductive inductive learning. Journal of Applied Logic 7, 3, 329 – 340.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	SEITZER, J. 1997. Stable ilp: Exploring the added expressivity of negation in the background knowledge.
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	SRINIVASAN, A., MUGGLETON, S., AND BAIN, M. 1996. Distinguishing exceptions from noise in non- monotonic learning,in s. muggleton and k. furukawa, editors, second international inductive logic pro- gramming workshop (ilp92).
Processed_A_New_Algorithm_to_Automate_Inductive_Learning_of_.txt	WU, X., KUMAR, V., ROSS QUINLAN, J., GHOSH, J., YANG, Q., MOTODA, H., MCLACHLAN, G. J., NG, A., LIU, B., YU, P. S., ZHOU, Z.-H., STEINBACH, M., HAND, D. J., AND STEINBERG, D. 2007. Top 10 algorithms in data mining. Knowl. Inf. Syst. 14, 1 (Dec.), 1–37.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	The best place to search for direct CP violation is the already observed charmless b → s modes. In SM with FSI, aCP in Kπ modes could be as large as 20–30% but diﬀer in sign between K −π+/K −π0 and ¯K 0π−. We illustrate possible New Physics eﬀects that could lead to aCP of order 40–60% in Kπ and φK modes distinguishable from FSI. b → sγ modes can also exhibit interesting asymmetries.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	so aCP > 10% in pure penguin modes would imply New Physics! We therefore have a discovery window in the next few years for beyond SM (BSM) eﬀects. The question then is: What BSM? Is large aCP possible in b s modes? Rather than trying to be exhaustive, we wish to demonstrate that CP asym- s transitions can indeed be large with simple extensions of SM, metries in b and sometimes even within SM.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	sg (with g on-shell) sg just does ∼ not give tangible signatures, our experimental knowledge of the strength of In fact, the long-standing “deﬁcit” in B decay c8 is actually rather poor. semileptonic branching ratio ( Bs.l.) and charm counting (nC ) point towards sg in Nature. 3,4 If b the possibility of sizable b 10%, which implies sg Bs.l. and nC can each be lowered by that amount and the problems c8 ∼ would go away. A recent CLEO bound 5 gives b sg < 7% at 90% C.L., D ¯DK + X decay. But even if which comes indirectly from the study of B sg at 1–5% would be one takes this seriously there is still much room, and b very hard to rule out. What we stress here is: if c8 is large in Nature, it must be coming from New Physics and should carry naturally a KM-independent CP violating phase.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	To have large b sγ constraint at the same time, one → needs additional source for radiating gluons but not photons. Gluinos (˜g) ﬁt the bill nicely. In SUSY one usually sets soft squark mass terms to be “universal” to suppress FCNC and to reduce the number of parameters. But it has been shown 4,9 that nonuniversal soft m ˜dj masses could give large c8 without violating the b In previous studies, however, the → possibility of new CP phases were not considered.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	The theory of inclusive decays are cleaner since one can use the quark/parton language. The absorptive parts arrising from short distance perturbative rescattering 12 can be used and one is insensitive to long distance phases. However, experimental detection poses a challenge, unless partial reconstruc- tion techniques can be made to work.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	s decays, one is interested in CP Since penguins dominate charmless b → s ¯dd and s¯ss. But since these violation in pure penguin processes such as b S), care 13 has to be taken in treating rates and asymmetries occur at CP violation in the b s¯uu mode, which has the distinction of receiving also the tree contribution. Although the tree amplitude alone does not lead to CP violation, while tree–penguin interference occurs only at (αS), to be consistent with treating pure penguin CP asymmetries, one needs to take into account the absorptive part carried by the gluon propagator (bubble graph) (α2 associated with the penguin. This S) tree–penguin interference term is needed to maintain CPT and unitarity in rates and hence aCP .
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	smallness. However, one power of αS is factored out by convention in deﬁning S) and its size is comparable to c1 within SM. One O8, hence c8 starts at has to keep track of the relevant leading order in αS when comparing with “full theory” approach discussed earlier.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	when interfering with the tree amplitude. We note that the use of operator language can be misleading at this stage since the absorptive parts are not resummed while the Wilson coeﬃcients are. One could easily lose track of αS counting that is needed for maintaining CPT/unitarity if one thinks too naively in eﬀective theory language.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	Having made all these precautions, we can square amplitudes in a straight- forward manner to obtain rates and arrive at the asymmetries. Since at lower 2 term has a log q2 pole behavior. We reg- order one has b c8\| s ¯dd ulate it by simply cutting it oﬀ at 1 GeV. Fig. 1(a) gives the rates for b (solid) and ¯b b. The SM result does not show a → prominent low q2 tail since b sg is small, and the asymmetry comes mostly from below c¯c threshold. For larger q2 the aCP is GIM suppressed. 13 The SM asymmetry is indeed tiny. For new physics enhanced c8 = 2eiσ, we consider the cases for σ = π/4, π/2 and 3π/4. Besides a very prominent low q2 tail since 10%, the salient feature is the rather large rate asymmetries b above c¯c threshold. The reason is because the new physics σ phase now evades the SM constraint of Eq. (1), and the c8 amplitude interferes with standard c3−6 penguins which carry the absorptive parts due to (perturbative) c¯c rescat- tering, but the u¯u rescattering absorptive part is suppressed by Vub. Note that for σ = π/4, 3π/4 one has constructive, destructive interference, respectively.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	b for (a) b → s ¯dd, (b) s¯uu and (c) s¯ss decays (solid) Figure 1: and ¯b decays (dashed). Curves with prominent small y tail are for c8 = 2eiσ with σ = π/4 (top), π/2 (middle), 3π/4 (bottom), while the other is SM result.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	Table 1: Inclusive Br (in 10−3)/aCP (in %) for SM and for c8 = 2eiσ.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	physics. Can one distiniguish between new physics eﬀects and SM with large FSI phases? The answer is yes, if one compares several modes.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	To illustrate the eﬀect of FSI, we now write A = A1/2 + A3/2eiδ and plot in Fig. 2(c) and (d) the Br and aCP vs. δ for γ = 64◦. The rate is not very sensitive to δ which reﬂects penguin dominance, but aCP can now reach 20%, even 30% for K −π0. We stress that the naively pure penguin ¯K 0π− mode is in fact also quite susceptible to FSI phases as it is the isospin partner of K −π0, which deﬁnitely receives tree contributions. The B− ¯K 0π− mode can re- ceive tree contributions through FSI rescattering. Comparing Fig. 2(b) and (d), aCP in this mode can be much larger than the naive factorization result. However, the aCP for ¯K 0π− and K −π+ are out of phase, hence, comparing the two cases can give information on the FSI phase δ.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	Figure 2: Br and aCP vs. (a), (b) SM unitarity angle γ, (c), (d) FSI phase δ for γ = 64◦, and (e), (f) new physics phase σ for γ = 64◦ and δ = 0. Solid, dotted, dashed and dotdashed lines are for K −π+, ¯K 0π−, K −π0 and φK respectively.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	is less sensitive to the FSI phase δ. In ﬁtting Br’s, we ﬁnd that destructive interference is necessary which can be understood from the inclusive results of Fig. 1. This means that large aCP s are preferred! We plot in Fig. 2(e) and (f) the Br and aCP vs. the new physics phase σ, for γ = 64◦ and δ = 0. The K −π+ and ¯K 0π− modes are very close in rate for σ 180◦, but the K −π0 mode remains a factor of 2 smaller. However, the aCP can now reach 50% for K −π+/ ¯K 0π− and 40% for K −π0! These are truly large asymmetries and would be easily observed, perhaps even before B Factories turn on (i.e. with CLEO II.V data!). They are in strong contrast to the SM case with FSI phase δ, Fig. 2(d), and can be distinguished.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	One can now construct an attractive picture. We have noted that recent φK upper limit within SM. If c8 is en- studies cannot explain the low B φK can be hanced by new physics and interferes destructively with SM, B K −π+ brought down to below 5 10−5 which is 2–3 times larger follows from c8 dominance, and their rate than the φK mode suggests a low ms value and slight tunings of BSW form 145◦, the rates are largely accounted for, but aCP for factors. Around σ φK, K −π+/K −π0 and ¯K 0π− could be enhanced to the dramatic values of 55%, 45% and 35%, respectively, and all of the same sign. This is certainly distinct from the sign correlations of SM with FSI.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	On the down side, within the scenario of strong penguin dominance, which includes the case of enhanced c8, the B− K −π0 rate is always about a factor of two smaller than the K −π+ mode, and we are unable to accommodate recent CLEO ﬁndings. 16,18 We are also barely able to accommodate B ωK. Within SM one needs 1/Neff ≃ 1.5 0 one can account for only half. Adding new physics induced c8 = 2eiσ eﬀect, we are able to account for Br for both large and small Neff, but not for N = 3. However, aCP is never more than a few % and hence not very interesting. Since the ωK mode has a single isospin amplitude, it is insensitive to FSI rescattering phases.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	K resonances. However, besides the B contributions to the inclusive spectrum has not yet been disentangled.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	sγ rate already, one has the Since SM accounts for the observed b c′ 7 to be small. constraint 7\| We ﬁnd that aCP larger than 10% is possible in certain parameter space, especially when the new physics eﬀect has opposite sign w.r.t. SM.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	We must recall that B physics has had its share of surprises. The long b lifetime was discovered without much theory encouragement. Bd mixing was in fact discovered with theory “discouragement”. More recently, the η′K and ωK modes turn out to be much larger than theory expectations, while the huge inclusive fast η′ + Xs simply came out of the blue without theory warnings. We therefore must be on guard for CP violation.
Processed_Large_and_Fun_CP_Violation_in_B_Meson_Decays:_Wher.txt	I thank Chun-Khiang Chua, Xiao-Gang He, Ben Tseng and Kwei-Chou Yang for collaborative work, and R.O.C. National Science Council for support.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Predicting ambulance demand accurately in ﬁne resolutions in space and time is critical for ambulance ﬂeet management and dy- namic deployment. Typical challenges include data sparsity at high resolutions and the need to respect complex urban spatial domains. To provide spatial density predictions for ambulance demand in Mel- bourne, Australia as it varies over hourly intervals, we propose a pre- dictive spatio-temporal kernel warping method. To predict for each hour, we build a kernel density estimator on a sparse set of the most similar data from relevant past time periods (labeled data), but warp these kernels to a larger set of past data irregardless of time periods (point cloud). The point cloud represents the spatial structure and geographical characteristics of Melbourne, including complex bound- aries, road networks, and neighborhoods. Borrowing from manifold learning, kernel warping is performed through a graph Laplacian of the point cloud and can be interpreted as a regularization towards, and a prior imposed, for spatial features. Kernel bandwidth and de- gree of warping are eﬃciently estimated via cross-validation, and can be made time- and/or location-speciﬁc. Our proposed model gives sig- niﬁcantly more accurate predictions compared to a current industry practice, an unwarped kernel density estimation, and a time-varying Gaussian mixture model.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Ambulance demand is often exceedingly sparse at the temporal and spatial resolution required for prediction. There is zero demand in the vast majority of 1-km2 regions over a 1-hour period.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	This demand arises from complex urban geography. The city boundary is often highly irregular. Ambulance demand can be very high (coastal and downtown) or very low (suburbs) along the boundary. Within this boundary, demand follows closely the city’s infrastructure and terrain; there might be high demand along central highways and zero demand within an internal lake. High resolutions covariates of these features are not readily available.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Ambulance demand exhibits spatial and temporal patterns. Weekly seasonality is usually prominent (Channouf et al., 2007; Matteson et al., 2011); the industry relies heavily on this seasonality to make pre- dictions. Some studies have also noted daily seasonality and short-term serial dependence at densely-populated regions (Zhou et al., 2015). • Ambulance demand data for large cities is often large-scale. This presents computational challenges, especially since predictions are needed very frequently.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	It is particularly diﬃcult to simultaneously resolve these challenges. Over- coming sparsity requires considerable smoothing, while capturing complex spatio-temporal patterns requires ﬁne-resolution modeling. At high granu- larities, data sparsity makes it diﬃcult to detect spatio-temporal character- istics accurately. At low granularities, diﬀerences across regions and times are not suﬃciently captured for optimal ambulance planning.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Figure 1 demonstrates these challenges in predicting ambulance demand for Melbourne. We show on the right the locations of 696, 975 demand inci- dents from years 2011 and 2012 (in gray), and those of 38 demand incidents for a typical 1-hour period (in black). On average, 99.6% of the 1-km2 re- gions in Melbourne receive zeros calls in any given hour. Comparing to a map of Melbourne on the left (Google Maps, 2015), we observe a highly com- plex spatial boundary as Melbourne encloses a large bay to its south-west. Demand is high near the bay, but low on the outskirt suburbs. Demand is visibly higher at small satellite suburban neighborhoods and along major highways radiating out from the city center. There is lack of demand due to several reservoirs and a national park to the west and northwest. Consistent with typical patterns, the demand exhibits strong weekly seasonality.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 1: Left: map of Melbourne (Google Maps, 2015); right: spatial locations of all 696, 975 Melbourne ambulance demand incidents from years 2011 - 2012 (in gray), and 38 demand incidents for a typical 1-hour period (in black). We observe complex boundary and geographical features (e.g., high- ways, roads, satellite suburbs).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	corresponding counts in the same hour of the preceding four weeks for the past ﬁve years (Setzler, Saydam and Park, 2009). Averaging so few historical counts, which are mostly zeros, produces noisy and ﬂickering predictions, resulting in haphazard and ineﬃcient deployment.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.