Bring back SSRL
jvmncs opened this issue · comments
jvmncs commented
The semi-supervised RL module was a naive attempt at an agent that can learn in misspecification problems with help from a supervisor. It was primarily used to help motivate library development, but it's become stale with recent changes. The goal of this issue is to reboot the semi-supervised learning demonstration agents. One possibility is to update the existing TabularSSQAgent.
Another, more ambitious opportunity could be to attempt applying the Spiky CRMDP regret bound to reduce the number of queries to the supervisor in cases where the Spiky CRMDP assumptions are weakly or strictly satisfied.