A possible route to parallel action-value computation with one prefill
A reply speculates that concatenating a prefix state embedding with different action embeddings could make a single prefill sufficient.
TLDR
A commenter suggests a possible setup: concatenate a prefix state embedding with different action embeddings. They speculate that this would allow prefill to run just once, with the values of all actions computed in parallel.
Combined views
67
1 Source, first seen ago