OpenAI Five bots beat a team of former pros at Dota 2

2 min read

Back in June, OpenAI Five, the artificial intelligence bot team, had smashed amateur humans in the video game Dota 2. But, this time it set completely different standard by beating semi-professional players at the Dota 2 game, yesterday. This was part of an effort to benchmark the progress of the bots so far as the openAI team plans to beat a team of top professionals at The International Dota 2 championship, taking place from August 20 to 25.

As mentioned in the OpenAI blog, “Dota 2 is one of the most popular and complex esports games in the world, with creative and motivated professionals who train year-round to earn part of Dota’s annual $40M prize pool (the largest of any esports game)”. It requires its players to have fast-twitch reflexes, strong knowledge of game strategies, along with solid teamwork.

OpenAI Five fight as a group and consists of five neural networks. The game started off with OpenAI Five playing warm-up games with the audience. Later, the OpenAI Five team played a three-game series, against a group of humans, that included former Dota 2 professionals and casters, Merlini, Fogged, Cap, and Blitz.

The OpenAI Five performed really well in the first two games. For the final game, the OpenAI team let the audience select their team of five heroes, which handicapped the bots’ chances of winning. Humans won the last round, finishing the series with a score of 2-1.

According to the OpenAI blog, “OpenAI Five plays 180 years worth of games against itself every day, learning via self-play. It trains using a scaled-up version of Proximal Policy Optimization running on 256 GPUs and 128,000 CPU cores”.

Proximal Policy Optimization ( PPO ) refers to a new class of reinforcement learning algorithms. It has become the default algorithm of choice for OpenAI as it is easy to implement and performs well. It involves “computing an update at each step that minimizes the cost function”.

The OpenAI team had made small changes to their neural network bot last month. This included increasing its reaction time and using new strategies.

This is not the first time when a computer has beaten human beings in games. From computers beating humans at chess to computers winning debates against humans, the chances of OpenAI Five beating professionals at The International does not seem that low.

But what has set apart OpenAI Five’s achievement from the rest is its ability to optimize strategies well in a team as opposed to simply learn to master strategies as individual players.

If you enjoyed reading this, be sure to check out the entire account on twitch, Dota 2 game: OpenAI Five vs Humans.