urbit
Minified exception occurred
That's right, it always uses the prediction. We thought that it might be interesting to switch between the two policies, but we found it doesn't really help. Maybe it is worth exploring further though!
Thanks for the reply. I have one more question, I run the evaluation and find that the evaluation can stop at a certain step (< 500 steps) for some unsuccessful episodes. How do you implement it?
Oh that's a normal failure case, it will happen if there is a false positive detection and the agent calls "stop" as it thinks it reached the target.
Yeah, I see.
And for the prediction model, seems the map size for training is 960 * 960, but during inference, the map size is 720*720.
Am I right?
Moreover, I find that only the target prediction inside the local map is used for update the global goal. Is it reasonable to train the prediction model only with the local maps?
Thanks for your patience.
Currently it is set that way, but maybe that is actually a mistake -- if you try 960 during inference, let me know what happens!
It is better to predict using global map than only using local map. See the ablation study in our paper for comparison
I thought the performance might not be significantly impacted. But actually, a 1% decrease in success rate on the validation set when using the map size 960, which is perplexing to me.
I believe only very few episodes need the agent to go outside of the 720 size map, so there is not much benefit to predicting out there. And then the new predictions causes some slightly different behavior sometimes which leads to 1% decrease in success (not really significant).