Harsh Agrawal, Shweta B. Guja

Abstract

OpenAI Five is a machine learning project developed by OpenAI that acts as a team of  game bots to play against humans in the competitive computer game Dota 2. In the game Dota 2, “The International” is the world championship tournament. A game between Dota 2 world champions and game bots developed by OpenAI was organised and the result was mind blowing. The International world champion ‘OG’ lost back to back games against OpenAI Five. OpenAI five observes the game after extracting the present game state from Dota developer’s API with one layer which contains 1024- unit LSTM. Without any human data involved, the neural network conducts actions via numerous available action heads and each head has meaning. In other words, bots play against each other learning best actions countering every possible move, in short opponent modelling. Opponent modeling is the strategy to observe and anticipate the moves and ability of an opponent. In gaming, the model is an abstracted description of the opponent and his strategy, countering the opponent's behavior in the game. This method is important in multiple agent settings where secondary agents with competing goals also adapt with their abilities [2], yet it remains challenging because strategies interact with each other and change. The result showed that AI (if trained in expert supervision) can be proved effective against a human mind in difficult situations.