Policy and Complex Systems
amount of error as the fixed initial conditions control across all four subnetworks in table 3 . The Jaccard distance values in table 4 , which can be read as a percentage , show a smaller difference , 1.6 percentage points different in the union subnetwork and less in the other three subnetworks . These initial results , with error rates only slightly worse than one type of control and better than another type of control , indicate that our method is plausible though the algorithm requires further refinement to improve its performance .
Comparing the results in tables 3 and 4 with the statistics in table 1 , error rates and network density appear correlated . The model runs with the highest error values also have the highest network densities . This pattern continues to hold between the highest and lowest values . There are too few values to confirm a correlation , but this result does suggest that our algorithm performs best on less dense networks .
Rates of Link Accretion and Link Decay
Phase 1 of our modeling analysis allows for a closer examination of the performance of algorithm for adding network links . Each execution of our algorithm , for accretion and for decay , can be expected to add additional error , so that higher rates of error are expected for the phase 2 analysis . These higher rates are seen , but the difference between control runs and algorithm runs is decreased . Seeing this convergence , it is worth asking how many links the accretion algorithm is adding and how many links the decay algorithm is removing . Table 5 records values for expected and observed values of both link accretion and decay , as well as a ratio of observed accretion and decay to
16