Learning Bayesian networks from data: An information-theory based approach

This paper provides algorithms that use an information-theoretic analysis to learn Bayesian network structures from data. Based on our three-phase learning framework, we develop efficient algorithms that can effectively learn Bayesian networks, requiring only polynomial numbers of conditional independence (CI) tests in typical cases. We provide precise conditions that specify when these algorithms are guaranteed to be correct as well as empirical evidence (from real world applications and simulation tests) that demonstrates that these systems work efficiently and reliably in practice.

A reactive force field simulation of liquid-liquid phase transitions in phosphorus

A force field model of phosphorus has been developed based on density functional (DF) computations and experimental results, covering low energy forms of local tetrahedral symmetry and more compact (simple cubic) structures that arise with increasing pressure. Rules tailored to DF data for the addition, deletion, and exchange of covalent bonds allow the system to adapt the bonding configuration to the thermodynamic state. Monte Carlo simulations in the N-P-T ensemble show that the molecular (P-4) liquid phase, stable at low pressure P and relatively low temperature T, transforms to a polymeric (gel) state on increasing either P or T. These phase changes are observed in recent experiments at similar thermodynamic conditions, as shown by the close agreement of computed and measured structure factors in the molecular and polymer phases. The polymeric phase obtained by increasing pressure has a dominant simple cubic character, while the polymer obtained by raising T at moderate pressure is tetrahedral. Comparison with DF results suggests that the latter is a semiconductor, while the cubic form is metallic. The simulations show that the T-induced polymerization is due to the entropy of the configuration of covalent bonds, as in the polymerization transition in sulfur. The transition observed with increasing P is the continuation at high T of the black P to arsenic (A17) structure observed in the solid state, and also corresponds to a semiconductor to metal transition. (C) 2004 American Institute of Physics.

Development of complex classical force fields through force matching to ab initio data: Application to a room-temperature ionic liquid

Recent experimental neutron diffraction data and ab initio molecular dynamics simulation of the ionic liquid dimethylimidazolium chloride ([dmim]Cl) have provided a structural description of the system at the molecular level. However, partial radial distribution functions calculated from the latter, when compared to previous classical simulation results, highlight some limitations in the structural description offered by force fieldbased simulations. With the availability of ab initio data it is possible to improve the classical description of [dmim]Cl by using the force matching approach, and the strategy for fitting complex force fields in their original functional form is discussed. A self-consistent optimization method for the generation of classical potentials of general functional form is presented and applied, and a force field that better reproduces the observed first principles forces is obtained. When used in simulation, it predicts structural data which reproduces more faithfully that observed in the ab initio studies. Some possible refinements to the technique, its application, and the general suitability of common potential energy functions used within many ionic liquid force fields are discussed.

The impact of virtual simulation in palliative radiotherapy for non-small-cell lung cancer.

Background and purpose: Radiotherapy is widely used to palliate local symptoms in non-small-cell lung cancer. Using conventional X-ray simulation, it is often difficult to accurately localize the extent of the tumour. We report a randomized, double blind trial comparing target localization with conventional and virtual simulation.Methods: Eighty-six patients underwent both conventional and virtual simulation. The conventional simulator films were compared with digitally reconstructed radiographs (DRRs) produced from the computed tomography (CT) data. The treatment fields defined by the clinicians using each modality were compared in terms of field area, position and the implications for target coverage.Results: Comparing fields defined by each study arm, there was a major mis-match in coverage between fields in 66.2% of cases, and a complete match in only 5.2% of cases. In 82.4% of cases, conventional simulator fields were larger (mean 24.5+/-5.1% (95% confidence interval)) than CT-localized fields, potentially contributing to a mean target under-coverage of 16.4+/-3.5% and normal tissue over-coverage of 25.4+/-4.2%.Conclusions: CT localization and virtual simulation allow more accurate definition of the target volume. This could enable a reduction in geographical misses, while also reducing treatment-related toxicity.

Optimising the use of virtual and conventional simulation: a clinical and economic analysis

Background and purpose: Currently, optimal use of virtual simulation for all treatment sites is not entirely clear. This study presents data to identify specific patient groups for whom conventional simulation may be completely eliminated and replaced by virtual simulation. Sampling and method: Two hundred and sixty patients were recruited from four treatment sites (head and neck, breast, pelvis, and thorax). Patients were randomly assigned to be treated using the usual treatment process involving conventional simulation, or a treatment process differing only in the replacement of conventional plan verification with virtual verification. Data were collected on set-up accuracy at verification, and the number of unsatisfactory verifications requiring a return to the conventional simulator. A micro-economic costing analysis was also undertaken, whereby data for each treatment process episode were also collected: number and grade of staff present, and the time for each treatment episode. Results: The study shows no statistically significant difference in the number of returns to the conventional simulator for each site and study arm. Image registration data show similar quality of verification for each study arm. The micro-costing data show no statistical difference between the virtual and conventional simulation processes. Conclusions: At our institution, virtual simulation including virtual verification for the sites investigated presents no disadvantage compared to conventional simulation.