How Did OpenAI's Dota 2 Bot Beat Dendi in a 1v1?

605.6K views
•
August 12, 2017
by
OpenAI
YouTube video player
How Did OpenAI's Dota 2 Bot Beat Dendi in a 1v1?

TL;DR

OpenAI's Shadow Fiend bot beat Dendi in the first game of a 1v1 mid match at The International 2017. The bot had no hard-coded strategy and did not learn from human experts. It learned by playing lifetimes of 1v1 games against copies of itself, starting from complete randomness and improving in small steps until it reached pro level.

Transcript

well as this final match is getting set up for the day in between the last matches we have a costume competition and now we have a little bit of a special treat before we get into that lower bracket match a special treat indeed sure and this this is something special so get your bum in those seats I see you you think it's time to get nachos it's no... Read More

Key Insights

  • OpenAI's Dota 2 bot learned entirely through self-play. It played lifetimes of 1v1 matches against a copy of itself, and the team only coached it on what they considered good or bad behavior, with no programmed understanding of the game.
  • The bot began training from complete randomness. Early games mostly ended with a Shadow Fiend dying to a tower, because the agents simply wandered around the map without any useful strategy or understanding of danger.
  • The bot's progress came from a series of small improvements rather than one breakthrough. It moved from hiding in base to pushing waves, then last hitting, then playing at a casual level, then aggressive play and finally baiting opponents.
  • The exhibition match used strict 1v1 rules: Shadow Fiend mirror in the middle lane, with no bottle, no runes, no neutral creeps, no raindrops, no shrines and no soul ring, and the match decided by kills.
  • The bot's advantage came from volume of experience. Because robots do not eat or sleep, it played Dota around the clock, explored far more of the strategy space than any human, and learned to exploit opponents.
  • Professional players described the bot as feeling like a smarter version of themselves. Several said it was very hard to kill, very fast and nearly impossible to beat, unlike the standard Dota bots they considered awful.
  • Dendi lost the first game and admitted he needed more patience. Before the second game he changed his approach and tried to copy the bot's habits, such as pushing the creep wave and constantly buying regeneration.
  • OpenAI's Dota project was built by a team of about ten developers. They brought the bot to The International to see how it compared with the best players in the world after it had already beaten other pros.

Explore YouTube Video Summarizer or Get YouTube Transcript Extractor

Questions & Answers

Q: How did OpenAI's Dota 2 bot learn to play?

OpenAI's Dota 2 bot learned through self-play. The team did not hard code any strategy and did not have it learn from human experts. Starting from complete randomness, it played lifetimes of 1v1 games against a copy of itself, and the team coached it on what they considered good or bad. Through many small improvements it eventually reached professional level.

Q: Did Dendi beat the OpenAI bot at The International 2017?

Dendi lost the first game against OpenAI's Shadow Fiend bot at The International 2017. After the loss he said "well played" and admitted he needed more patience. Before the second game he decided to try a different approach, pushing the creep wave and copying what the bot was doing. The bot had already beaten other professional players during the week before the match.

Q: What were the rules of the OpenAI vs Dendi 1v1 match?

The match was a Shadow Fiend mirror in the middle lane. The rules were no bottle, no runes, no neutral creeps, no raindrops, no shrines and no soul ring. The match was decided by kills, so it became a pure test of laning, last hitting, creep control and trading damage between two players using the same hero.

Q: What did the OpenAI bot do in the early stages of training?

In the early stages of training the bot behaved completely randomly. Most games ended with one of the Shadow Fiends dying to a tower because the agents just ran around the map. They then figured out a safer strategy of staying in base, occasionally venturing out to push the creep wave a little, and over time learned to last hit.

Q: Was there one big breakthrough in OpenAI's Dota bot training?

According to OpenAI's Jakob, there was no single big breakthrough. Progress came from a series of small improvements. The bot first reached the level of a casual player, then learned to play aggressively, and then learned to bait opponents. Each step built on the last until it was able to compete with and beat professional players.

Q: Why could the OpenAI bot beat professional Dota players?

The bot beat professionals because it accumulated far more experience than any human. It played lifetimes of Dota games, explored many different strategies and learned to exploit opponents. As the host summarized, robots do not eat or sleep, so the bot played Dota 24 hours a day and covered much more of the strategy space than any human player.

Q: How did professional players describe playing against the OpenAI bot?

Professional players who tested the bot said it was nothing like the normal Dota bots, which they called pretty awful. They described it as feeling like playing a smarter version of themselves, said it was very fast and nearly impossible to kill, and several expressed shock that a bot could be better than they had ever imagined.

Q: What strategy did Dendi try in his second game against the bot?

For the second game Dendi planned to push the creep wave, explaining that 1v1 differs from 5v5 because no one can come from behind and gank you, so pushing to get levels faster is safer. He said he was trying to copy what the bot was doing, including constantly buying regeneration, which he called the cheapest item exchange in the matchup.

Summary & Key Takeaways

  • At The International 2017 in Key Arena, Dendi, a player who had attended every International, faced an OpenAI bot in a 1v1 Shadow Fiend mid match. The rules banned bottle, runes, neutral creeps, raindrops, shrines and soul ring, making it a pure laning contest decided by kills.

  • OpenAI's Greg explained that the bot was not programmed with Dota knowledge. It played lifetimes of 1v1 against itself and was coached on good and bad behavior. Pros who tested it beforehand found it fast, hard to kill and better than they had imagined possible for a bot.

  • Dendi lost the first game after the bot outplayed him in creep control and trades. OpenAI's Jakob described training stages from randomness to pro play. For game two, Dendi planned to push waves, buy regeneration constantly and copy the bot's own habits.


Read in Other Languages (beta)

Share This Summary 📚

Explore More Summaries from OpenAI 📚