The proposed preference attention layer does not seem necessary about preferencetransformer HOT 2 CLOSED

Hello!

1. I believe I may have missed some details in the README. It is correct to set the "--transformer.use_weighted_sum" flag to True.

2. The evaluation performance is highly sensitive to the seed number, particularly in the Mujoco locomotion environments. For our experiments, we utilized seed numbers starting from 3407. However, it's possible that the code may not be able to reproduce the results in the paper precisely due to various factors such as the version of JAX/Tensorflow libraries, type of GPUs, etc. The gain from using the preference attention layer is comparatively marginal in Mujoco locomotion environments as the task and reward function is relatively Markovian when compared to Antmaze navigation tasks.

from preferencetransformer.

Hi, I additionally ran the Antmaze experiments since you suggested the performance gain would be more pronounced in Antmaze:

It still seems the performance is better off without the preference attention layer.

I understand that the evaluation performance of RL can be sensitive to random seeds (or other subtle implementation details), but I am doubtful whether the preference attention layer consistently having no performance gain on all of the 8 tasks can solely be attributed to those factors.

Thank you for the response.

from preferencetransformer.