Last week, Open AI’s new algorithm that is trained to play the complex strategy game, Dota 2, beat the world champion e-sports team OG at an event in San Francisco, winning the first two matches of the ‘best-of-three’ series. The competition included a human team of five professional Dota 2 players and AI team of five OpenAI bots.
OpenAI Five makes history by winning a best-of-three versus @OGesports, the Dota 2 world champions!!!! Huge congrats to the team and to OG for an extremely well-played match.
— Greg Brockman (@gdb) April 13, 2019
I usually like to refer to Five as analogous to an insect — trained on an evolutionary timescale, operates via ingrained “instincts” which handle even unexpected things in its environment, but can’t quickly adapt to pick up new skills.
Feels like a step forward but much remains. https://t.co/bK2Jjk42eA
— Greg Brockman (@gdb) April 14, 2019
Dota 2 is a complex multiplayer strategy game, where two teams of five players each, compete to destroy a large structure that is defended by the opposing team known as the “Ancient”, while also defending their own. The game consists of over 100 unique characters, deep skill trees, etc.
The five Open AI bots had been trained using deep reinforcement learning techniques and had played 45000 years worth of Dota 2 gameplay in a span of 10 months. OpenAI constantly kept improving by playing itself in a virtual environment. During Dota 2, each team had to pick out of 17 heroes. In the first match, OpenAI used different aggressive tactics and human players managed to survive for more than 40 minutes. One such tactic used by the OpenAI bots involved spending the earned in-game currency to revive heroes upon death, during the start of the match.
In the case of the second match, OpenAI’s performance got even better and gained an early edge against OG. Overall, it took less than 20 minutes for the AI to win against humans. The last round was won by the humans.
This was the last planned public demonstration of OpenAI’s bot, however, Open AI’s research team is now working on new software that allows humans to collaborate alongside the OpenAI Five software in real time. This will allow the team to learn from its unique strategies and patterns.
OpenAI is also planning to release a platform, called Arena, that would allow the public to play against OpenAI Five. The platform will be open for three days from April 18th at 9 PM Eastern time. However, this collaboration may not be made available to the public.
Apart from that Sam Altman, Co-founder, and CEO, OpenAI, is focussing on a different area of research, which involves making the AI adapt to work in less perfect simulators. Also, OpenAI is working on using reinforcement learning to provide robotic hands more human-like movement, reports the Verge.
“What OpenAI is trying to do is build general artificial intelligence and to share those benefits with the world and make sure it’s safe. We’re not here to beat video games, as fun as that is. We’re here to uncover secrets along the path the AGI”, said Altman.
Public reaction to the news is largely positive with people congratulating OpenAI for the win:
Guys, before we go any further we just want to say we come in peace and we never doubted our new AI overlords to begin with 👍👍👍
— LGD Gaming (@LGDgaming) April 14, 2019
Huge congrats. Well deserved and inspiring to other researchers.
— grovetheory (@jackgrove2) April 13, 2019
That was amazing, huge improvement since the last live event. Five played incredibly well and it did not look weird, just good plays all around. Congrats to the team!
— Felipe Leite (@felipempleite_) April 13, 2019
OpenAI is rightly proud of this milestone. I’d hesitate to conclude much about artificial general intelligence, though.
Bounded games on which, as here, the system can accrue 45,000 years (!) of play against itself to train are quite different from acting out in the world. https://t.co/Ex1WvSlWug
— Jonathan Zittrain (@zittrain) April 13, 2019
Read Next
OpenAI LP, a new “capped-profit” company to accelerate AGI research and attract top AI talent
OpenAI introduces Neural MMO, a multiagent game environment for reinforcement learning agents