vis_encode_type for List<float> and creating channeled observations?

Unity Engine
, ,
1

For doing reinforcement learning on a board game (which has specific connected fields), I’d like to include some kind of convolutions for my network. I saw that you can specify this by selecting an appropiate vis_encode_type in the training config file, but I’m wondering if this is only going to work for visual observations (e.g. Camera/Texture sensors)?

I find it easier to encode the board state myself and feel like using a camera sensor would be way too much here, perhaps even misleading.

So the question is: can I still use convolutions for simple observations (List of floats)?

Additionally, I’d like to encode multiple channels (one for each type of game piece), how can I do this in the CollectObservations?

2

One option is to create your own custom sensor. This can then use ObservationWriter.Item[int32, int32, int32], see Class ObservationWriter | ML Agents | 2.3.0-exp.3

You could add your custom sensor to the board object, and it would return the state of the board.

3

(Note that I’m pretty sure that mlagents expects images to be at least 20x20)

4

Allright, I’ll have a look at that, thank you!
Can I still specify the vis_encode_type in the config, or do I have to do something specific on the C# side for that?