test_extracted_captions/kamrahHowDiverseInitial2023.json

{"CAPTION FIG1.png": "'Fig. 1 Correlation matrix showing the relative correlation between two gammas by comparing the way our DPP approach ranks 10,000 sampled sets of cardinality, \\\\(k\\\\!=\\\\!10\\\\). The gammas values in both axes here are logarithmic values with base 10.\\n\\n'", "CAPTION FIG3.png": "'\\n\\n## References\\n\\n* [1] A. A. Barabasi, A.\\n\\n'", "CAPTION FIG2.png": "'Figure 2: Scatter plots showing randomly chosen sets with \\\\(k=5\\\\) high-diversity and low-diversity samples with their diversity score on top of each of the chosen set\\n\\n'", "CAPTION FIG4.png": "'\\n\\n## 4 Conclusion\\n\\nFigure 4: Experiment 1: optimality gap grid plot showing the difference in current optimality gap between optimizers initialized with 5th versus 95th percentile diverse sample (y-axis) as a function of optimization iteration (x-axis). The different factors in the factor grid plot the effects of diversity as the noise amplitude and smoothness are varied in the range [0.2, 0.8]. Each plot also has text indicating the net cumulative optimality gap (NCOG), a positive value corresponds to a better performance by high-diversity samples compared to the low-diversity samples. The plot shows that BO benefits from diversity in some cases but not others. There is no obvious trends in how the NCOG values change in the grid. The results are further discussed in Sec. 3.\\n\\n'", "CAPTION FIG5.png": "'\\n\\n## References\\n\\nFig. 5: Experiment 2: box plot showing distribution of \u201clengthscale\u201d hyper-parameter learned by BO when initiated with diverse (right-side) and non-diverse samples (left-side) for 16 different families of wildcat wells functions of the same parameters but 100 different seeds. The optimal hyper-parameter for each of the 100 wildcat wells instances from each family is also plotted as horizontal lines\u2014in many but not all cases these overlap. Each cell in the plot also has the 95th percentile confidence bound on MAE for both diverse and non-diverse samples. The results show that MAE confidence bounds for non-diverse samples are smaller compared to diverse samples for all the families of wildcat wells function. Thus, indicating a presence of model building advantage for non-diverse initial samples. The results of this figure are further discussed in Sec. 4.\\n\\n'", "CAPTION FIG6.png": "'\\n\\n## 4 Conclusion\\n\\nFigure 6: Experiment 3: optimality gap plot showing effects of diversity when the optimizer is not allowed to fit the hyper-parameters for the Gaussian process and the hyper-parameters are instead fixed to the values found in experiment 2. The results from this plot show positive NCOG values for all families of wildcat wells function, showing that once the model building advantage is taken away, the diverse samples outperform non-diverse samples. Further discussion on this plot can be read in Sec. 5.\\n\\n'"}