These are fantastic tips. It's been awhile since I did any agent work but it's great to see others continuing experiments and coming up with best practices. Thanks for the video.
Hey there. Fantastic tips! Im planning to conduct a ML project on unity, and was wondering on the training speed i could achieve. On an environment like the one you created, how many instances could you get running simultaneouly? At 1:00 you seem to have 8 instances, but have you stress tested this limit? Also, does the timescale value you use (i remember seeing 20x in your other video) affect the training performance, or does it make no difference at all? Thanks!
Thanks for your comments and questions. I had not really stress tested the system in the video much beyond what is shown. So yes, there were 8 instances of the environment, but I had not tried more than that. I think you are right about training at 20x, but I had not tried other values to see how the results would compare. There are so many permutations of things that could be tried! Sorry not to have better answers. Good luck on your project!
We need to load models that we download in runtime from a webserver, and we execute it in a build (not in Editor), is it possible? All the examples and code I find only work within Editor using the Resources Folder and the walk scene only works in the editor and we i try to build it, it reverts back to the default ONNX Model. and not my trained model.
There is a function called SetModel() which is supposed to change the model at runtime. I have not used it much, but it is used in the WallJump example environment. I have a use for SetModel() in mind but haven't yet gotten to the point of using it. If you learn any more about it, I would be glad to hear about it.
These are fantastic tips. It's been awhile since I did any agent work but it's great to see others continuing experiments and coming up with best practices. Thanks for the video.
Thanks for the feedback!
Congrats on 700 subscribers 🔥
Thank you!
Hey there. Fantastic tips!
Im planning to conduct a ML project on unity, and was wondering on the training speed i could achieve.
On an environment like the one you created, how many instances could you get running simultaneouly? At 1:00 you seem to have 8 instances, but have you stress tested this limit?
Also, does the timescale value you use (i remember seeing 20x in your other video) affect the training performance, or does it make no difference at all?
Thanks!
Thanks for your comments and questions. I had not really stress tested the system in the video much beyond what is shown. So yes, there were 8 instances of the environment, but I had not tried more than that. I think you are right about training at 20x, but I had not tried other values to see how the results would compare. There are so many permutations of things that could be tried! Sorry not to have better answers. Good luck on your project!
Love the tips. Would love to see different config experiments and results from different rewards.
Thanks for the comment! I've done a lot of experiments with different rewards and need to make another video soon!
Great vid!
Thanks for the comment!
Very nice tips. I like your ml-agents videos.
Thank you!
We need to load models that we download in runtime from a webserver, and we execute it in a build (not in Editor), is it possible?
All the examples and code I find only work within Editor using the Resources Folder and the walk scene only works in the editor and we i try to build it, it reverts back to the default ONNX Model. and not my trained model.
There is a function called SetModel() which is supposed to change the model at runtime. I have not used it much, but it is used in the WallJump example environment. I have a use for SetModel() in mind but haven't yet gotten to the point of using it. If you learn any more about it, I would be glad to hear about it.
Should of made the dog an agent as well that tries to avoid being caught, thanks for these tips
Yes, one step at a time. :) And I do think there is a role for techniques besides ML; NavMesh is very nice.