A joint group of researchers from University College London and the Chinese Alibaba Group presented the BiCNet neural network, which has learned to masterfully play StarCraft.
The multifactor bidirectional coordinated network (Multiagent Bidirectionally-Coordinated Network, BiCNet) first mastered basic actions like moving units, and then moved on to more complex actions.
BiCNet discovered the tactics of “hit and run” and coordinated attacks, picked up effective combinations of units, learned how to analyze the actions of the opponent and use the landscape to gain an advantage.
Hit and run tactics
Coordinated attack with cover
In combination with the speed of microcontrol inaccessible to humans, BiCNet began to play StarCraft at the level of progamers.
The creators of BiCNet claim that their algorithm has the potential to solve large-scale problems not only in games, but also in the real world.
Source: Cornel University