Improbable Icon


ANN with SpatialOS?



Did anyone ever tried training neural nets using SpatialOS? I’m curious if Spatial is suited for that kind of projects.

If it has not been done, anyone else thinking about it?


Hello @SionoiS,

I know of at least two people who have tried out machine learning with SpatialOS, I am however not sure whether they are on the forum (EDIT: actually… CC @j.e). The one example I am most familiar with consists of a custom C++ worker that uses TensorFlow for the ML part.

That said it really depends on what kind of neural net you want to train. Some use-cases do not really benefit from the particular features that SpatialOS offers (e.g training a chess AI) whereas others might see their ability to learn efficiently increase ten-fold because SpatialOS provides a better setting to create the right learning scenario’s.

Happy to discuss the topic as I am actually pretty interested in this subject myself for potential personal projects. Currently I am reading and learning about (deep) neural nets and doing the research to see what might be possible and in what domains SpatialOS could be best put to use.


I did a little project using ANNs on SpatialOS. I implemented deepmind’s asynchronous actor-critic method based on this paper:

The basic idea is that each entity trains based on its own experience, and then periodically sends and receives gradients with a shared central network.
In my brief experience, it worked pretty well. I was able to get simple agents to learn to interact with a basic environment and share their experience with each other. I’d love to see it pushed further.


@dvanamst I am learning about all kind of neural nets too! I’m still not sure what I want to do but it seem that SpatialOS would be a better fit for experimenting with ANN topology and other network science related stuff.

@chris Very cool! It definitely take advantage of the way SpatialOS works.

I have this feeling that I know what kind of experiment would work well with SpatialOS but I’m lacking words right now.

Thanks for your inputs guys.


I definitely will not hesitate to post anything down here as soon as some I have more concrete ideas / designs / results so we can keep the discussion going. :slight_smile: