Google DeepMind Reflects on 15 Years of AI Research Driven by Games

Games as a Foundation for AI Research

Google DeepMind has utilized games as a primary driver for AI research since its inception in 2010. These constrained yet rich environments have been instrumental in understanding intelligence and developing significant AI breakthroughs. The company's founders and team members have backgrounds in game development, fostering a deep respect for the craft.

Early Breakthroughs with Atari

The journey began with a small team training a deep neural network to play Atari 2600 games directly from raw pixels. This led to the development of the Deep Q-Network (DQN), which learned to play 49 different games without specific engineering. The 2015 Nature paper on DQN helped establish the modern era of deep reinforcement learning.

Mastering Complex Games

Following Atari, DeepMind tackled more complex games, leading to more capable and general AI systems. AlphaGo defeated world champion Go player Lee Sedol in 2016. Subsequent iterations, AlphaGo Zero and AlphaZero, learned from self-play and generalized to master chess, shogi, and Go with a single algorithm. In 2019, AlphaStar achieved Grandmaster level in StarCraft II, demonstrating proficiency in real-time strategy with imperfect information.

Impact on Gaming and AI

These AI advancements have not only pushed the boundaries of artificial intelligence but also enriched the gaming experience. AlphaGo's 'Move 37' in Go and AlphaZero's new lines of play in chess inspired experts and altered established strategies. DeepMind continues to partner with game developers, including Fenris Creations and the EVE Universe, to prototype new gameplay experiences that advance both gaming and AI.