Groundhog Day

Kenton White | December 22, 2020  | Article | 2 Comments

“What would you do if you were stuck in one place, and everything that you did was the same, and nothing mattered?” — Phil Connors

This is the question posed by Bill Murray’s character in the movie Groundhog Day.  Weatherman Phil Connors must relive his worst day ever in the Pennsylvania town of Punxsutawney and home of Punxsutawney Phil, a groundhog who, if he sees his shadow, foretells six more weeks of winter.  Phil is doomed to repeat his mistakes until he wins the heart of producer Rita Hanson (Andie MacDowell).  While hypothetical for us, this question — what would you do — is very relevant for Polly, who can keep repeating an election until she gets the right result.

So what would Polly do?  In the movie, Phil relives day after day, learning the right response for every possibility the day may throw at him.  What poem should Phil recite to Rita?  What toast should he make?  What should he say each step of the way?  Despite repeating the same day over and over and over again, he is unable to cover all the possibilities and fails at winning Rita’s heart.

Many people think this is how Polly learns.  That she repeats the same election with different rules until she gets the result right.  Underestimate Miami Cubans?  Program Polly to account for this.  Miss Trump-supporting Hispanics in Texas?  Program Polly to find more of this group.  Yet, like with Phil, scripting every outcome, it doesn’t work.  In Artificial Intelligence, this is called overfitting — crafting the code to get perfect agreement with one outcome.  Problem is, when something new comes up this kind of AI can’t adapt.  With the 2020 US election it was Cubans in Florida.  Next election it is Amish in Pennsylvania.  

So how does Phil eventually win Rita’s heart?  In the movie, he stops focusing on Rita and starts focusing on other people in Punxsutawney.  Rather than winning the heart of a single person, he focuses on winning the hearts of the town.  By learning from hundreds of interactions instead of a single one, Phil knows just the right response for any situation.  He can adapt to changing situations, using what he has learned to win anyone’s heart, including Rita’s, and ending his loop.

Just like Phil, this is how Polly learns.  If Polly focuses just on Miami Cubans, this is like Phil focusing just on Rita.  Instead, Polly has been reviewing data from dozens of elections where minority groups had an outsized influence.  This includes the Wildrose party in Alberta, the Bloc Quebecois party in Quebec and the Green party in British Columbia.  Through a process called transfer learning, Polly does what Phil did in Groundhog Day — she generalizes from each of these experiences so that she can adapt.

We’ve had great success so far.  She is better able to forecast regions she missed, like Victoria and Vancouver Island in BC — two strong showings for the Green Party in the last Provincial election – and of course Florida and Texas in the recent US election.  But the real test will be when, like Phil, Polly applies this general learning to new elections.  At the end of Groundhog Day, Phil used what he learned from the townspeople, winning Rita’s heart in the end.  Polly, too, learns from many examples and that helps to ensure the next election isn’t Groundhog Day all over again.

About these results

Submit a comment

Your email address will not be published. Required fields are marked *

2 thoughts on “Groundhog Day

  1. I think all, in all Polly did a good job predicting the American election. It did state that Biden would win both the electoral, and popular votes, which was true. It equally predicted that Michigan, Pennsylvania, and other big swing states would go to Biden, and they did. North Carolina, Florida, and Texas were the exceptions. Is Polly looking at the run-off senate races in Georgia? Any trends so far?

    1. Thanks for the vote of confidence!

      Polly won’t be forecasting the Georgie runoff election. Here is the answer from Shainen Davidson, Senior AI Scientist:

      “The quality of available polling isn’t sufficient to train on directly. I looked a bit into using transfer learning from other elections as training, but because this one is a special election it doesn’t map well onto other election data.”

Comments are closed.

Polly’s day at the races

asiadmin | February 12, 2021


Liberal Roller Coaster

asiadmin | August 26, 2021


Share This