Last August at the Dota 2 International Tournament (Ti8) in Vancouver a team of OpenAI bots was twice defeated by human professional players. After eight months of development efforts, the “OpenAI Five” exacted their revenge today against one of the world’s top teams in a highly anticipated best-of-three 5v5 Dota 2 showdown in San Francisco.
OpenAI’s bot system pulled out a tough victory in the first game and crushed the International 2018 Champions “OG” in the second game in less than 20 minutes. Said OG player Johan Sundstein, nicknamed BigDaddy, in his post-match interview: “They are really great. Once you are sucked into their games, that is when you never win.”
OpenAI was founded in San Francisco in 2015 with the ultimate goal of building an Artificial General Intelligence (AGI) capable of performing a multitude of tasks within one general system. Researchers have found a useful testing ground in virtual real-time environments such as those in Dota 2, one of the world’s most popular multiplayer online battle arena (MOBA) video games. OpenAI’s first Dota 2 bot system made an impressive debut at the Dota 2 International 2017 in Seattle, defeating top Ukrainian gamer and Dota 2 pro “Dendi” in a one-on-one match.
OpenAI then demonstrated a big step forward by developing a five-bot coordinated AI team. The OpenAI Five dispatched a team of casters and ex-pros with MMR rankings in the 99.95th percentile of Dota 2 players last August, but failed in the next two games against Brazilian pro team “paiN” and five Chinese Dota 2 legends.
Today’s OpenAI opponent was Europe-based professional esports team and The International 2018 Champions “OG.” The team scored 68 percent in the Luckbox online poll, “Who would you love to see win the battle of man vs machine – @OGesports or @OpenAI?”
The OpenAI research team built its current bots’ brains with Long Short Term Memory (LSTM), a unit in a Recurrent Neural Network (RNN) proficient at remembering information for long periods of time and well-suited to classifying, processing and making predictions based on time series data. They also used other techniques such as exponential decay factor to determine whether the team should look at long-term rewards or short-term rewards, and a new “team spirit” hyper-parameter to improve its in-game collaboration.”
OpenAI Researcher Filip Wolski told Synced that they increased the size of the current bot system compared to the Ti8 versions, from 40 million parameters to 160 million parameters, which he likened to roughly ten times the brain of a fruit-fly and one-tenth the brain of a honeybee. The training time of the system was also longer, from the previous 10,000 human years of gameplay to 45,000 human years.
Dota 2 is a highly complex and wildly popular video game played between two teams of five players. The team that takes down their opponent’s center base “Ancient” wins the game. The game environment has 115 characters and all-important “Heroes,” 22 defensive towers, dozens of non-player characters, hundreds of skills and items, and a long tail of game features such as runes, trees, wards, and so on.
Today’s game rules were similar to those for the Ti8 competition: certain items (divine rapier, bottle) were eliminated; the number of heroes was reduced from 18 to 17; techniques such as scams and illusions were forbidden; and the number of couriers was one.
First game: Open AI 1 – 0 OG
- OG: Viper, EarthShaker, Witch Doctor, Stealth Assassin, Shadow Fiend
- OpenAI Five: Sven, Sniper, Crystal Maiden, Death Prophet, Gyrocopter
The OpenAI bots and OG players exchanged the smallest bit of “trash talk” before the contest, the bots boasting their estimated winning probability had topped 80 percent, OG deadpanning “okay” in their text reply. OpenAI Five drew first blood in the first team fight.
After some ups and downs, OG made a costly mistake about 20 minutes in, losing their tier-two tower and uplands. At 32 minutes OpenAI’s bots won a key team fight and pressured OG’s base. GG. Synced was told that compared to the bots that played last year at Ti8, the new versions’ improvements include team fights and warding, an important skill which enables vision in unknown areas. It was observed that bots still have little idea of how to use couriers to deliver items, and choose bad times to revive heroes using “buyback.”
Second game: Open AI 2 – 0 OG
- OG: Sniper, EarthShaker, Death Prophet, Lion, Slark
- OpenAI Five: Crystal Maiden, Witch Doctor, Viper, Sven, Gyrocopter
OpenAI’s bots quickly built advantages by better farming and team fights. In 18 minutes the bots had earned 25,000 more golds than OG and gained 31 more kills, a huge gap rarely seen in professional Dota 2 games. The game quickly went into trash time, and OG surrendered at 21 minutes, surprising even OpenAI: “The first game was a bit longer, so I thought there would be more back and forth in the second game,” OpenAI Researcher Jonathan Raiman told Synced.
The event wrapped up with a performance between joint teams comprising two human players and three OpenAI bots, in which a William “Blitz” Lee Team scored a narrow win. This Synced reporter was offered a chance to engage the bots in a demo, and in 12 minutes of gameplay failed to get one kill, foiled by the bots’ impressive farming, lane control and zoning abilities.
Great news for Dota 2 otaku is that OpenAI has decided to open up an online battle arena from April 18 to 21, giving humans everywhere the chance to challenge the bots either in competitor mode or in teammate mode. Registration will open on April 13.
Founded as a non-profit, OpenAI recently announced it was incorporating a new for-profit arm into its organization. While the surprising move has raised some questions regarding future focus and development, OpenAI Science Communicator Ashley Pilipiszyn told Synced their Dota 2 project has not been affected. She also said that although OpenAI has not identified any specific business plans built on this research, “We are looking at things like creative systems. There are a lot of possible applications.”
As Synced has previously reported, the major contribution of OpenAI’s Dota 2 project is using relatively simple techniques to enable complex coordination and long horizon gameplay in an imperfect game environment. “We prove that reinforcement learning doesn’t have to be narrow,” said Raiman.
The OpenAI Five victory ends the company’s long journey toward Dota 2 mastery, marking another AI milestone two years after DeepMind’s historic victory over Go Master Ke Jie. Although countless challenges remain, OpenAI has to believe today’s breakthrough also advances them along the road to AGI.
Journalist: Tony Peng | Editor: Michael Sarazen