Born to Do Science: January 2015

Last Sunday we presented the first of the current series of four Born To Do Science programs a the Stillwater Public Library. My guest was Allan Axelrod, who studies machine learning.

Allan has developed an algorithm (or set of computer instructions) he calls EIEIO, which helps computers get clever about how they collect and analyze data in situations where numbers are collected continuously, but the entire set cannot be seen at once.

The situation that brought about the need for this algorithm is this: How do we monitor carbon sequestration sites for potentially dangerous leaks of CO2 back into the atmosphere?

We began with a song inspired by Allan's whimsical EIEIO acronym.

Introducing the topic with a song.

We had some discussion about whether carbon sequestration (storing waste carbon dioxide under ground) is a good idea in the first place. I know it's a controversial topic! But we needed to put a pin in that, in order to get to Allan's algorithm. Given that CO2 is already being sequestered under ground, it's good that people like Allan are coming up with ways to monitor it!

We set up a game to mimic the situation Allan has to deal with. There are monitoring stations on the ground spread over a large and hard-to-get-to area of land. Each one takes CO2 measurements every hour. The computer can use this information to model the system and attempts to predict leaks before they occur.

The problem is that all the current CO2 data is not available instantly. Drones are used to fly by each station and pick up the data. The question is - should the drones visit every station once, then every station again, etc.? Or is there a more useful way of gathering the data?

One of our sampling stations "collecting" data.

For the game, we set up seven sampling stations around the room. Two volunteers became drones, and two others were the computer. The computer sent drones from station to station collecting data. The computer then tried to make good decisions about where to send them next based on the numbers collected.

(If you're wondering - I created a set of measurements for each station ahead of time. The station-masters flipped through a stack of cards slowly, one every ten seconds. Each card they put down represented taking a measurement. At each station, the drones collected all the measurements taken since the last visit.)

The object of the game was to find a leak (if the numbers went high enough, they turned red on the page representing a leak) as quickly as possible.

A "drone" picks up data from a sampling station.

The "computer" analyzes data, trying to decide where to send the drones next.

Processing...

The kids tried a couple of basic strategies. One was to send drones randomly to different stations and hope to find the leak. Another was to notice if the numbers were going up from a given station and send drones back there, or to a nearby station. It was challenging to process all those numbers during the game. But sending drones to a station where the numbers were rising is very similar to Allan's actual strategy!

If I'd had the chance to test the game out first, I would have changed a few things about how we did it to make it easier and clearer for the kids to think about strategy. But we did successfully get a handle on the basics of how the station, drones, and computers work together in real life.

Allan answers questions.

Allan's strategy with the EIEIO algorithm is to compare data collected from each station to the numbers the computer would have predicted. A "metric" - a measure of how close the numbers are to the predicted value - is calculated for each station. The stations that are the farthest off get the next drone visits.

This way the computer can collect the numbers it most needs next at any given time.

Allan has tried several versions of his algorithm on several sample data sets and compared the results of his EIEIO strategy to the strategy of sending drones methodically or randomly to all the stations every time. His EIEIO results are much better!