har-early-decision: code for "Learning when to stop: a subject-level, cross-platform evaluation of reinforcement-learned early decisions for inertial human activity recognition"
Year: 2026
Type: software
Abstract: Code for the paper "Learning when to stop: a subject-level, cross-platform evaluation of reinforcement-learned early decisions for inertial human activity recognition" (G. Çetin, submitted to Measurement Science and Technology, 2026). A causal CNN-GRU prefix classifier produces class probabilities after every 8-sample chunk of an inertial window; a stopping policy trained with PPO on subject-level cross-fitted classifier outputs decides when to stop. The package contains the complete pipeline (causal preprocessing, cross-fitting, PPO, all comparators, subject-level statistics), all analyses added during revision, unit tests, reproduction checks, and the scripts that generate all numbers, tables and figures of the manuscript from the result files. The data sets (UCI HAR, WISDM, OPPORTUNITY, PAMAP2) are not redistributed; see README.md. The result files are deposited in a separate record.
Author: Çetin, Gökhan
Institution: Gümüşhane University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225751341