David Lin

I'm still very much in the process of constructing this site. In the meantime, I invite you to read about some of the more interesting projects that I'm currently working on.

Almost three years ago, when me and two of my college friends (Lester Mackey and David Weiss) decided to enter the Netflix Prize Competition, we had no idea what we were getting ourselves into. This was a machine-learning competition aimed at improving Netflix's movie recommender system (with the winner getting $1 million!).

It all started as a fun side project, and one of the first things our team had to do was decide on a name. After a couple hours of brainstorming, we thought, why not use the first movie in the dataset. The movie turned out to be Dinosaur Planet. (It was a close call between that or the last movie in the dataset: Alien Hunter...!) After a few more hours of debate, we decided that dinosaurs were far cooler than aliens. And so, Dinosaur Planet had been formed!

After three dramatic years (more details to come in the future), here we are, now part of a grand coalition that is currently sitting atop the leaderboard (only ahead of the second play team by the most narrow of margins).

The path to getting here was long and eventful, but the short version goes like this. Dinosaur Planet merged with team Gravity to form When Gravity and Dinosaurs Unite to take a shot at the one year progress prize. We ended up losing to BellKor (AT&T labs) by a very small margin. This team then founded the Grand Prize Team, to try to amass a larger coalition. Soon after, GPT would merge with Vandelay Industries (another big coalition) and Opera Solution to form The Ensemble!

Now whether Dinosaur Planet, which started as a fun spontaneous side project, can be part of a coalition to win the biggest machine-learning competition in the world...I sure hope so!

In early July of this year, the University of Alberta ran the 2009 Computer Poker Competition. There are lots similarities to the Netflix Competition, especially in the need to write highly efficient algorithms and data structures. One variant of poker, limit head's up, is quickly become a game where computers are now stronger than the best human players.

My bot, Rockhopper, had a pretty good showing, finishing 4th out of 13 competitors in the equilibrium competition (where you're scored based on how many other bots you can beat). It beat one of the two Hyberborean bots (by University of Alberta), and lost in its match against the second place bot by an almost insignificant margin.

The University of Bucharest's bot (GGValuta) was definitely the strongest one overall, and their description (using k-means to cluster nodes in later stages of the hand) has given me a lot of ideas for next year's competition.

You can see the full match results of the competition here.