You can repair damage to the policy from RL by merging the base weights back in. Why not do simulated annealing with the annealed variable being how much interpolation you do with the base weights…

Posters: jd_pressman. jd_pressman

2 likes recorded by the source corpus.