Improvement Equivalence for Decision-Aware Model-Based Reinforcement Learning
Formalize and test when model-based RL models can discard observational detail yet preserve useful policy improvement. The project treats models as joint experience generators P(s,a,r,gamma,s') and studies the hierarchy OE <= VE <= GE <=… Read moreRead less
Formalize and test when model-based RL models can discard observational detail yet preserve useful policy improvement. The project treats models as joint experience generators P(s,a,r,gamma,s') and studies the hierarchy OE <= VE <= GE <= IE with explicit policy, state, action, value-function, and parameter quantifiers.