Agent needs to be able to plan long-term as well as react and adjust to counter strategies. It demands full capabilities of macro economy, tech and upgrade, army compositions, battle tactics, attack and defense timing, micro control of units, etc.
The decision-making action space of Go is within 361, while StarCraft action space is as many as 10 to the 26th power. The complexity grows exponentially with the growth of multi-dimensional decision factors such as decision timing, instructions, subjects, objects, etc.
Different from game such as Go which is complete information, StarCraft has a "fog of war" mechanism. The map and opponent information is not available without scouting and opponent modeling. Different from Texas poker which is also incomplete information game, StarCraft environment is dynamic as the battlefield changes rapidly.
StarCraft requires the coordination of many units (up to two hundreds), whereas games such as "DOTA2" and "Glory of the King" only require control of five units.
In order to be competitive in StarCraft, the agent needs to make decisions in Millisecond level. This imposes a challenge on performance and efficiency of the neural network model and optimization.
Large scale neural network models, for processing of spatial, sequential and unit information, auto-regressive, multi-task learning, sequence to sequence modeling, reasoning under incomplete information.
Research on cutting edge distributed reinforcement learning framework to support large-scale parallel training, efficient training algorithms, exploration and self-play mechanism to improve sample efficiency.
Research on multi-agent learning and evolution learning, improve agent robustness and diversity.
For business inquiries, please fill in the contact form and we will get in touch as soon as poosible.