Incorporating game mechanics into daily tasks has proven to be an effective way to motivate workers. As it turns out, gamification techniques don’t just work on us. Google DeepMind is applying the tactic to machine learning.
Prodded by gamification techniques, artificial intelligent (AI) systems are quickly becoming game masters: There’s IBM Watson and Jeopardy!, Google DeepMind and Go, and Carnegie Mellon University’s Libratus and No Limit Texas Hold ‘Em Poker champion — the latter a landmark victory in the annals of AI gaming because poker involves bluffing and guesswork.
The triumphs don’t stop there. AI is also becoming a video game master. The Google DeepMind team has trained computational AI systems known as neural nets to play Atari video games such Breakout, which was released in 1976.
The objective of Breakout, a single-player game, is to rid the top third of the video screen of bricks. But “the machine was not given the rules of Breakout,” Erik Brynjolfsson, professor of management at the MIT Sloan School of Management and director at the MIT Center for Digital Business, said the recent MIT Disruption Timeline conference.
Instead, the machine was given the raw pixels of the screen; a controller, which moves left and right; and an objective to maximize the score. After 500 games, the neural nets performed better than humans, even developing new strategies, Brynjolfsson said.
Here’s the real punchline: Researchers at DeepMind then took the process of training neural nets on how to win at Atari video games and turned it into a gamification technique for energy efficiency. Researchers trained a system of neural nets on operating scenarios, historical data on energy consumption as well as prediction data and gave it access to all of the gauges and dials; this time, the objective of “the game” was to maximize energy efficiency, a huge cost center for the internet search giant.
“Now, this data center had already been heavily optimized by a bunch of very smart PhDs, some of the best in the world,” Brynjolfsson said. “So this is not an easy problem at all.”
Turns out, the neural nets bested the best, managing a 15% reduction in overall power savings and a 40% reduction of energy used for cooling, one of the biggest consumer of energy, in particular.
“You can imagine if you take that level of improvement and apply it to all of our systems — our factories, our warehouses, our transportation systems, we could get a lot of improvement in our living standards,” Brynjolfsson said.