Offline meta-reinforcement learning (OMRL) utilizes pre-collected offline datasets to enhance the agent's generalization ability on unseen tasks.

A/B testing is critical for modern technological companies to evaluate the effectiveness of newly developed products against standard baselines. This paper studies optimal designs that aim to maximize the amount of information obtained from online experiments to estimate treatment effects accurately.

While this is not close to the state-of-the-art in general (c.f. Figure 2 shows the effect of action entropy on exploratory behavior in Montezuma's Revenge. Number of unique avatar positions visited. Full training curves across all 6 Atari games are shown in Figure 1, including the random policy baseline. To ensure this didn't hamper performance, we At each state visited by the agent evaluator during training, the agent's state (consisting of the avatar's The full curves are included for completeness. The compute cluster we performed experiments on is heterogenous, and has features such as host-sharing, adaptive load-balancing, etc.

In fact, the bound can be improved by considering a data-dependent measure, e.g., Rademacher

In offline RL, a critical challenge is distribution shift (also called "extrapolation error" in literature).

A widely used technique for improving policies is success conditioning, in which one collects trajectories, identifies those that achieve a desired outcome, and updates the policy to imitate the actions taken along successful trajectories. This principle appears under many names -- rejection sampling with SFT, goal-conditioned RL, Decision Transformers -- yet what optimization problem it solves, if any, has remained unclear. We prove that success conditioning exactly solves a trust-region optimization problem, maximizing policy improvement subject to a $χ^2$ divergence constraint whose radius is determined automatically by the data. This yields an identity: relative policy improvement, the magnitude of policy change, and a quantity we call action-influence -- measuring how random variation in action choices affects success rates -- are exactly equal at every state. Success conditioning thus emerges as a conservative improvement operator. Exact success conditioning cannot degrade performance or induce dangerous distribution shift, but when it fails, it does so observably, by hardly changing the policy at all. We apply our theory to the common practice of return thresholding, showing this can amplify improvement, but at the cost of potential misalignment with the true objective.

Random datasets are gathered by a random policy.