15 research outputs found
Learning an Ensemble of Semantic Parsers for Building Dialog-Based Natural Language Interfaces
No full textUsing Multiple Clause . . .
No full textIn this paper, we explored a learning approach which combines different learning methods in inductive logic programming (ILP) to allow a learner to produce more expressive hypotheses than that of each individual learner. Such a learning approach may be useful when the performance of the task depends on solving a large amount of classification problems and each has its own characteristics which may or may not fit a particular learning method. The task of semantic parser acquisition in two dierent domains was attempted and preliminary results demonstrated that such an approach is promising
Scaling Up ILP to Large Examples: Results on Link Discovery for Counter-Terrorism
No full textInductive Logic Programming (ILP) has been shown to be a viable approach to many problems in multi-relational data mining (e.g
Cluster Analysis of Simulated GravitationalWave Triggers Using S-MEANS and Constrained Validation Clustering
No full textThe fifth Science run of LIGO (S5) has been concluded recently. The data collected over two years of the run calls for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. The study presents two new techniques for cluster analysis of gravitational wave burst triggers. Traditional approaches to clustering treats the problem as an optimization problem in an “open” search space of clustering models. However, this can lead to problems with producing models that over-fit or under-fit the data as the search is stuck on local minima. The new algorithms tackle local minima by putting constraints in the search process. S-MEANS looks at similarity statistics of burst triggers and builds up clusters that have the advantage of avoiding local minima. Constrained Validation clustering tackles the problem by constraining the search in the space of clustering models that are “non-splittable” models in which centroids of the left and right child of a cluster (after splitting) are nearest to each other; the region of models that either over-fit or under-fit data (i.e. “splittable” models) can therefore be effectively avoided when assumptions about data are satisfied. These methods are demonstrated by using simulated data. The results on simulated data are promising and the methods are expected to be useful for LIGO S5 data analysis
An experimental comparison of genetic programming and inductive logic programming on learning recursive list functions
No full textThis paper experimentally compares three approaches to program induction: inductive logic programming (ILP), genetic programming (GP), and genetic logic programming (GLP) (a variant of GP for inducing Prolog programs). Each of these methods was used to induce four simple, recursive, list-manipulation functions. The results indicate that ILP is the most likely to induce a correct program from small sets of random examples, while GP is generally less accurate. GLP performs the worst, and is rarely able to induce a correct program. Interpretations of these results in terms of differences in search methods and inductive biases are presented
S-means: Similarity Driven Clustering and Its application in Gravitational-Wave Astronomy Data Mining
No full textClustering is to classify unlabeled data into groups. It has been well researched for decades in many disciplines. Clustering in massive amount of astronomical data generated by multi-sensor networks has become an emerging new challenge; assumptions in many existing clustering algorithms are often violated in these domains. For example, K means implicitly assumes that underlying distribution of data is Gaussian. Such an assumption is not necessarily observed in astronomical data. Another problem is the determination of K, which is hard to decide when prior knowledge is lacking. While there has been work done on discovering the proper value for K given only the data, most existing works, such as X-means, G-means and PG-means, assume that the model is a mixture of Gaussians in one way or another. In this paper, we present a similarity-driven clustering approach for tackling large scale clustering problem. A similarity threshold T is used to constrain the search space of possible clustering models such that only those satisfying the threshold are accepted. This forces the search to: 1) explicitly avoid getting stuck in local minima, and hence the quality of models learned has a meaningful lower bound, and 2) discover a proper value for K as new clusters have to be formed if merging them into existing ones will violate the constraint given by the threshold. Experimental results on the UCI KDD archive and realistic simulated data generated for the Laser Interferometer Gravitational Wave Observatory (LIGO) suggest that such an approach is promising
Search for gravitational waves from low mass compact binary coalescence in 186 days of LIGO’s fifth science run
No full textWe report on a search for gravitational waves from coalescing compact binaries, of total mass between 2 and 35M⊙, using LIGO observations between November 14, 2006 and May 18, 2007. No gravitational-wave signals were detected. We report upper limits on the rate of compact binary coalescence as a function of total mass. The LIGO cumulative 90%-confidence rate upper limits of the binary coalescence of neutron stars, black holes and black hole-neutron star systems are 1.4×10−2, 7.3×10−4 and 3.6×10−3 yr−1 L−110, respectively, where L10 is 1010 times the blue solar luminosity
Search for long-lived gravitational-wave transients coincident with long gamma-ray bursts
No full textLong gamma-ray bursts (GRBs) have been linked to extreme core-collapse supernovae from massive stars. Gravitational waves (GW) offer a probe of the physics behind long GRBs. We investigate models of long-lived (~10–1000 s) GW emission associated with the accretion disk of a collapsed star or with its protoneutron star remnant. Using data from LIGO’s fifth science run, and GRB triggers from the Swift experiment, we perform a search for unmodeled long-lived GW transients. Finding no evidence of GW emission, we place 90% confidence-level upper limits on the GW fluence at Earth from long GRBs for three waveforms inspired by a model of GWs from accretion disk instabilities. These limits range from FtoFcm-2, depending on the GRB and on the model, allowing us to probe optimistic scenarios of GW production out to distances as far as = 33 Mpc. Advanced detectors are expected to achieve strain sensitivities 10x better than initial LIGO, potentially allowing us to probe the engines of the nearest long GRBs
