Considering the success in of Gradient boosting machines in machine learning, one might think of the application of the same principle to elections. The rationale behind it being that voters may have changed their vote conditioned on the knowledge of the current power distribution is. Unconditional voting may lead to power distributions which may be unexpected or unwanted by majority of voters. For example, the referendum in favour of the BREXIT may have caused voters a posteriori to decide otherwise (“If I only had known, I would have voted otherwise”).

A potential procedure which is feasible in the age of high performance computing and electronic communication would be: The percentage of 100% is equally distributed across all parties. A first voter has a a budget of percentage of 5%, to change the power distribution across all eligible parties according to his preferred distribution. The second voter again has a budget of 5% to distribute his budget in favour of his/her preferred distribution. This process is continued until the distribution reaches a steady state or all voters have had a chance to apply their voting budget. Since later voters will have a larger impact/more knowledge of what the final power distribution will be and therefore will have more influence, one might reduce their voting budget in a stepwise manner to compensate for that and reduce the influence of single voters on the final power distribution. The next voter may be selected based on largest deviation from the current power distribution, not in terms of absolute percentages, but in relative changes, ie changes of votes of smaller parties will count more than changes in votes of larger parties. The assumption being that more extreme deviations from the current power distribution indicate more knowledge about the best power distribution.

The algorithm is unsupervised in the sense that the optimal distribution of power is unknown, but supervised in the sense that we assume that each voter has an idea about the optimal distribution of power.

The opposite idea of such an conditional demographic election would be a single “expert” who decides on the optimal distribution of political power among parties. The obvious disadvantage of this approach would be bias by personal interests.

Considering the success of gradient boosting machines to model outcomes one might assume that such an conditional demographic election approach may reflect the ideal/truely desired distribution of political power better than the current unconditional demographic votes.