OpenAI Dota 2 Bots Defeat Human Players Using Team Neural Networks
OpenAI trains a team of five neural networks to defeat human opponents in the real-time game Dota 2. The system plays 180 years' worth of games daily to prepare for an upcoming esports championship.
OpenAI creates a team of five neural networks, called OpenAI Five, that successfully defeats skilled human opponents in the complex video game Dota 2. Unlike previous turn-based AI challenges like chess or Go, this system requires rapid real-time decision-making at 30 frames per second over 45-minute matches.
The AI system learns entirely by playing against itself, processing the equivalent of 180 years of Dota 2 games every single day. Running on 124,000 cores on Google Cloud, this massive computational effort represents a significant leap from OpenAI's earlier 1v1 experiments to coordinated team play.
Despite current restrictions on certain characters and strategies, OpenAI plans to fully test its system at a Dota 2 esports world championship in August. Before that major event, the organization hosts a Twitch-streamed tournament next month to publicly showcase the capabilities of its advanced gaming AI.