Limited by game speed

StarCraft II API
First of all, thank you Blizzard and Deepmind starting this amazing project. I'm a long time Starcraft player and a ML researcher.

I was wondering if the SC2 API available to us would be limited by the game speed, or that time is arbitrary and CPU speed will limit the amount of steps you can perform per time unit.

Secondly, will there be any type of relationship with Tensorflow? Something like a python module that can interface with Tensorflow and SC2 python API, or Deepmind's own research team utilizing Tensorflow?
11/05/2016 11:01 AMPosted by GrackyChan
First of all, thank you Blizzard and Deepmind starting this amazing project. I'm a long time Starcraft player and a ML researcher.

I was wondering if the SC2 API available to us would be limited by the game speed, or that time is arbitrary and CPU speed will limit the amount of steps you can perform per time unit.

Secondly, will there be any type of relationship with Tensorflow? Something like a python module that can interface with Tensorflow and SC2 python API, or Deepmind's own research team utilizing Tensorflow?


You should be able to interface with Tensorflow using the Protocol Buffer interface they mention in the original post.

https://developers.google.com/protocol-buffers/
11/08/2016 11:18 AMPosted by Paul Keet
You will not be locked to real time playing. The API will allow you to run the game as fast as your computer can handle, or very slow if your agent needs the processing time.
11/10/2016 10:48 AMPosted by AngryBacon
11/05/2016 11:01 AMPosted by GrackyChan
Secondly, will there be any type of relationship with Tensorflow? Something like a python module that can interface with Tensorflow and SC2 python API, or Deepmind's own research team utilizing Tensorflow?


You should be able to interface with Tensorflow using the Protocol Buffer interface they mention in the original post.

https://developers.google.com/protocol-buffers/

There are both Python and C++ libraries for TensorFlow. Looks like Blizzard is planning a C++ library slated for Q1 2017 and a TBD Python library.
11/26/2016 08:57 PMPosted by Sixen
11/10/2016 10:48 AMPosted by AngryBacon
...

You should be able to interface with Tensorflow using the Protocol Buffer interface they mention in the original post.

https://developers.google.com/protocol-buffers/

There are both Python and C++ libraries for TensorFlow. Looks like Blizzard is planning a C++ library slated for Q1 2017 and a TBD Python library.


Currently (last time I checked anyway) the C++ TensorFlow library only lets you run graphs. You still need to construct the graph in Python. Depending on how you want to thread your algorithms, you may be better off keeping your AI code in a separate process anyway.

Join the Conversation

Return to Forum