Schrodinger for Prez? Quantum models may solve inaccurate election predictions

Data analysts show quantum approaches can avoid computing problems that bedevil election forecasting. Lauren Fuge reports.

Hillary Clinton at Donald Trump's Inauguration Lunch in January 2017, perhaps trying to work out how a 90% probability of winning still ended in defeat.
Hillary Clinton at Donald Trump's Inauguration Lunch in January 2017, perhaps trying to work out how a 90% probability of winning still ended in defeat.

Quantum computation can be used to train models to accurately predict election results, say US researchers.

The 2016 US presidential election result came as a shock to everyone who had been watching the opinion polls — many of which predicted that Hillary Clinton had a 90% probability of winning.

Post-election analyses revealed that the majority of forecasting models missed the mark because predictions from individual states did not correlate with each other, and so failed to capture the more complex behaviour of the country as a whole.

Integrating the various state-based results into a country-wide model that would express more complex, higher-order relationships, says a team of computer scientists led by Maxwell Henderson of Washington DC data analysts QxBranch, is theoretically possible, but is defeated by “roadblocks” thrown up by the practicalities of classical computing.

Henderson and his colleagues, however, may have found an alternative solution to the problem. In a paper published on preprint server on arXiv, the analysts reveal that quantum computation can be used to improve election forecasting. The researchers focus on using the method to train a tool called a Boltzmann machine to perform image recognition tasks —sampling from fully-connected graphs with arbitrary correlations.

A Boltzmann machine is a type of learning model known as a deep neural network. It comprises a number of simple computational units connected in a way analogous to the way neurons interact in the brain. Each unit makes randomised decisions about whether to be on off. The result is an algorithm that can discover complex patterns within large datasets.

The machines produce immensely powerful models, and can be trained to make probability distributions that reveal sophisticated insights. Their downside, however, is that in classical computing models they require vast amounts of input data, learn slowly, and are expensive to train.

Henderson and his team, however, show that quantum computing methods can be used to train Boltzmann machines using much less data, based on the system sampling from the individual state-based models.

To test their approach, the researchers let their machine loose on the state results from the Trump-Clinton contest. The predictions that emerged, they report, were equal to those achieved by the “best in class” classical models.

And although demonstrating that quantum computing models can miss the mark as accurately as the best classical ones is perhaps a dubious triumph, it nevertheless points to interesting possibilities for predicting the next presidential race.

“While quantum computers and samplers are an emerging technology, we believe this application area could be of near-term interest,” the researchers write. They conclude that their methods “could bring an interesting new factor into the conversation of election forecasting at large, one in which quantum computation could play a future role”.

Lauren Fuge is an Adelaide-based author and science communicator.
Latest Stories
MoreMore Articles