fluid owl
I'm trying ML for the first time for a racing game, but my agent stutters (blinks left/right rotations and on off acceleration) when I use the trained model and disable training.
I see in training my fps drops to ~8 (~160 when not training)
Not sure what info to share, please ask for any details that can help 🙂
gif in training
gif after training
What I call "after training" is when I disable my python backend and reference the .onnx updated during training in my agent > behavior parameters > model
stray sleet
During training mode, the timescale of your scene is set to 20×, this means that your simulation is running 20 times faster than inference mode (what you call "after training").
The stuttering happens in both cases, you just can't see it during training because it's running so fast it's less noticeable, it's still there regardless.
One way to do this is go to the Decision Requester component on you agent and increase the Decision Period (with take actions between decisions set to true).
Another way is to tweak the rewards.
fluid owl
During training mode, the timescale of your scene is set to 20×, this means that...
Thanks a lot for the answer, so am I supposed to train my models with minimal Decision Period then increase it when using inference mode (test multiple values until one seems correct?) My issue being that for a racing game I would want to keep it as low as possible since they should move pretty fast.
By tweaking the rewards I guess you mean rewarding a less shaky driving?
Last thing is there a way to change how many cycles of training I can input? My models always stop at 500k which seems limiting to me.