A research laboratory co-founded by Elon Musk, OpenAI, has designed a new kind of AI bots that are able to play Dota 2 (a complicated tactical game) in 5-on-5 multiplayer games. The new agents from OpenAI have so far been capable of beating semi-professional and amateur teams. With this achievement, the research lab is now seeking to convey its agents to The International this August. The International is a productive Dota 2 tournament.
The new agents are dubbed as OpenAI Five, an indication to the amount of neural networks operating together in the group. To teach the neural networks, the AI has been playing almost 180 Years worth of gameplay each day employing reinforcement education. This allows the AI to study the details of the game, taking into consideration that it is far more complex than board games such as Go and Chess. Dota 2, for instance, comprises hiding information from users, avoiding the system from getting the complete playing field at any particular time.
The hardware used by the research team to teach OpenAI Five is remarkable. The 5 neural networks are taught via an enhanced version of Proximal Policy Optimization operating on 128,000 CPU cores and 256 GPUs. The similar setup was used last year in a much smaller scale when OpenAI launched out an AI system that proved competent of defeating the world’s best Dota 2 players in 1-on-1 games.
Speaking of AI, scientists at Stanford have designed a new AI program that organized the elements periodic table in just a couple of hours. This is a feat that took mankind almost a time frame of century and error. Dubbed as Atom2Vec, the program productively learned to differentiate between various atoms after analyzing a series of names for chemical compound from an online database, as per the resources.