Network induction for epidemic profiles with a novel representation

Graphs can be used as contact networks in models of epidemic spread. Most research seeks to extract the properties of an extant graph, derived from questionnaires or other sources of contact information. The inverse problem of searching the space of graphs for those that exhibit specific properties has received little attention and that is the focus of this study. This is, in part, because searching the space of contact networks is difficult. This paper extends and tests a representation for searching the space of contact networks with evolutionary computation. The focus of this study is on improvements in the representation used to evolve potential contact networks, adding an operator that permits strictly local adjustments to connectivity of the network, and another that does nothing at all. The benefits of doing nothing at some points during the construction of a network are substantial, because this permits evolution to adjust the number of active commands issued automatically. Adjusting local connectivity was identified as a beneficial feature in earlier research. The network induction method is tested on two tasks; finding a network that sustains an epidemic as long as possible and finding a network that, under simulation, closely matches a specified pattern of rise and fall in the number of infections.