The Untold Story of the Atari Benchmark | Approximately Correct Podcast

When Amii Fellow and Canada CIFAR AI Chair Mike Bowling and his colleagues started using old Atari video games from the 1980s to test his reinforcement learning (RL) agents, he didn’t know it would eventually become a benchmark used by researchers worldwide.

“It was pretty exciting to see other people see what I saw in it. But originally, I wasn’t really trying to change RL research. I just wanted to do RL research with this.”

On the first episode of Approximately Correct, Bowling dives deeper into what drew him to explore using 40-year-old video games to test leading-edge artificial intelligence, he talks about the technical hurdles that had to be overcome, how it opened up new avenues for testing and comparing new approaches to artificial intelligence, and how a small scientific curiosity expanded into a benchmark now used by research teams around the world. Bowling says the benchmark gives AI researchers a tool to quickly experiment and try out new ideas in machine learning research.

“We want to try out lots of different ideas. And if you can come up with a result on 50 days in a day or a week, instead of training for a year, then we can try out a lot of ideas.”

Bowling tells the fascinating untold story behind the Atari benchmark on the launch episode of Approximately Correct: An AI podcast from Amii. Hosted by Alona Fyshe, Amii Fellow, Canada CIFAR AI Chair, and Scott Lilwall, Amii’s Scientific Communicator, Approximately Correct delivers stories from the leading edge of AI research with compelling, rarely heard stories from Amii’s team of world-class researchers, leaders and thinkers. The podcast invites listeners to go beyond the buzzwords and hear the stories behind the science, meet the people advancing technology and learn about AI’s potential to transform our world from the people on the ground.

You can hear the inaugural episode of Approximately Correct on Spotify, Apple Podcasts, Google Podcasts and other podcasting services.