| [ | |
| "Theorem 3.5 proves that under Assumptions 3.1-3.3, the iterates of Algorithm 1 (Committed Q-learning) converge almost surely to Q*, with the resulting greedy policy optimal whenever the environment satisfies π-rewire-robustness (Theorem 3.5, Definition 3.7).", | |
| "Rewire-robustness (Definition 3.7) is proven to be strictly weaker than the q*-realizability condition (Definition 3.4) used in prior work, admitting a broader class of environments (Section 3).", | |
| "Committed Q-learning differs from standard Q-learning by committing to a single sampled action upon entering a feature and only resampling when the observed feature changes, rather than resampling every step (Section 3, Algorithm 1).", | |
| "Quasi-Markov environments are defined via an entrance matrix Σ satisfying p̄₀=ΣΦp̄₀, providing a structural condition analyzed alongside rewire-robustness (Definition 4.1, Section 4).", | |
| "In a corridor environment experiment, classical Q-learning fails to converge to the optimal reactive policy while Committed Q-learning converges quickly, illustrating the theoretical gap between the two algorithms (Section 4)." | |
| ] | |