ReinPatch

Summary

ReinPatch is a learned adaptive patching method for time-series forecasting. It trains a standalone patch-boundary policy with reinforcement learning and uses the downstream forecasting loss as the reward signal.

Role In The Wiki

ReinPatch is the current time-series case study for dynamic tokenization inspired by H-Net-like learned chunking. Its value is not that it becomes a full forecasting foundation model, but that it shows how a detachable patcher can be trained, frozen, and transferred across forecasting datasets.

Interface

  • Input unit: numeric time-series observations.
  • Patch decision: binary or multi-level boundary labels over the input sequence.
  • Optimization: Group Relative Policy Gradient over sampled boundary sets.
  • Downstream role: compress the temporal context before a forecasting backbone processes it.
  • Transfer mode: freeze a pretrained foundation patcher and train only the downstream backbone.

Evidence