batch reinforcement learning

Batch reinforcement learning is the older name for offline RL, from a time when the contrast was with online, incremental updates. You gather a batch of logged transitions once, then run your learning algorithm over that fixed set — often repeatedly — rather than interleaving learning with fresh data collection.

Classic batch methods like fitted Q-iteration treat value learning as a sequence of supervised regression problems on the batch: fit the value function to the current Bellman targets, recompute the targets, refit, and repeat. The data never changes between iterations; only the targets do.

In modern usage batch RL and offline RL are nearly synonymous, though offline RL emphasizes the hard distributional-shift problems that arise precisely because the batch is fixed and the learned policy keeps drifting away from it.