Dota 2 isn't the only game AI is learning to master. Google's DeepMind research teams have successfully used reinforcement learning to teach AI agents to play Quake III Arena Capture the Flag with "human-level" performance not only competitively, but also on teams with human players.
DeepMind's research works to push the boundaries of AI learning, working toward programs that could solve complex problems relating to climate change, healthcare, and other world issues. To do so, the teams have used games as a "training ground" of sorts to teach programs how to solve much smaller problems and now, work collaboratively to do so.
Though the idea behind the reinforcement learning programs is similar to the way OpenAI taught machines to play Dota 2 in teams, DeepMind's focus is on diversity in both team structure and opponents. AI learned Quake III Arena's Capture the Flag by playing matches against humans and other AI.
"Billions of people inhabit the planet, each with their own individual goals and actions, but still capable of coming together through teams, organisations and societies in impressive displays of collective intelligence," reads the official blog post announcing the research findings. "This is a setting we call multi-agent learning: many individual agents must act independently, yet learn to interact and cooperate with other agents. This is an immensely difficult problem - because with co-adapting agents the world is constantly changing."
Through multi-agent learning, DeepMind's AI learned to play Quake III Arena at the same level as skilled humans, even while playing alongside them in a tournament that included 40 human players randomly matched with AI as teammates and opponents. The AI eventually exceeded the winrate of the strongest human players in the group. It also began, like the Dota 2 AI, to demonstrate strategies popular with humans, such as opponent base camping.
"In the future, we also want to further improve on our current reinforcement learning and population based training methods," the post concluded. "In general, we think this work highlights the potential of multi-agent training to advance the development of artificial intelligence: exploiting the natural curriculum provided by multi-agent training, and forcing the development of robust agents that can even team up with humans."