The City of Philadelphia recently redesigned districts for its council members based upon the 2010 census. The districting process evinced considerable public interest and engagement because council districts from the prior census had significant shortcomings. During the 2010 redistricting process, several public interest groups came together to organize a districting contest. The organizers hoped to increase public engagement in the districting process and to proactively offer several well-constructed examples of city council districts that minimize gerrymandering. We were active participants in the contest, developing methodologies for finding good solutions to large integer programs that enabled us to win in one contest category and make presentations before the city council in this and another category. This article describes the unfolding of various events surrounding the districting process, the methodologies we developed, and the influence that the contest ultimately had on the design of city council districts.
Use of optimization models in science and policy applications is often problematic because the best available models are very inaccurate representations of the originating problems. Such is the case with electoral districting models, for which there exist no generally accepted measures of compactness, in spite of many proposals and much analytical study. This article reports on an experimental investigation of subjective judgments of compactness for electoral districts. The experiment draws on a unique database of 116 distinct, legally valid districting plans for the Philadelphia City Council, discovered with evolutionary computation. Subjects in the experiment displayed, in the aggregate, remarkable agreement with several standard measures of compactness, thus providing warrant for use of these measures that has heretofore been unavailable. The exercise also lends support to the underlying methodology on display here, which proposes to use models based on subjective judgments in combination with algorithms that find multiple solutions in order to support application of optimization models in contexts in which they are only very approximate representations.