haidang2405/senseshift-rl-real-world-active-sensing

SenseShift-RL

SenseShift-RL is a replayable active-sensing world built from real, quality-screened U.S. EPA ozone measurements. On each day an agent activates at most two of eight monitoring sites, observes only those valid measurements, and is rewarded for reconstructing the withheld network while paying an explicit authored sensing cost.

This is an observation-control benchmark, not an air-quality control simulator. Actions reveal recorded measurements; they do not alter pollution, weather, monitor operation, public policy, or health outcomes.

Version 1.1 contains 32 non-overlapping geographic worlds, 461,728 measurement-grid rows, and 426,221 valid recorded values. Worlds are frozen as 20 train, 6 calibration, and 6 final test worlds. The release includes a Gymnasium environment, non-learned and neural baselines, locked ablations, per-world effects, uncertainty analyses, model packages, and a workshop paper.

Load a table

Each logical table is exposed as its own subset. The default is measurements.

from datasets import load_dataset

measurements = load_dataset(
    "haidang2405/senseshift-rl-real-world-active-sensing",
    "measurements",
    split="data",
)
worlds = load_dataset(
    "haidang2405/senseshift-rl-real-world-active-sensing",
    "worlds",
    split="data",
)

Install the exact bundled environment implementation when you need the RL API:

pip install release/senseshiftrl-0.2.1-py3-none-any.whl

Schema and action semantics are documented in release/DATA_DICTIONARY.md, release/METHOD_CARD.md, and release/RESEARCH_PROTOCOL.md.

Integrity and scientific status

The directory release/ is the byte-identical audited v1.1 research release. Verify it with:

python release/verify_release.py --bundle release --mode full --json

Expected release SHA-256: 235b320ec4c9e495abf8fa40227da5fa706ef682ead4faad09b85967c347689a. The verifier checks 43 manifest-bound files and 102,546,600 bytes.

The confirmatory results and their interpretation boundaries are in release/RESULTS.md. In particular, Gaussian information—not deep RL—has the strongest observed deployable mean on the joint future-time-plus-unseen-geography track. The release preserves negative and non-significant results rather than claiming general RL superiority.

Source, rights, and responsible use

Measurements are derived from the exact AQShift-US release dee1eccb05f3..., sourced from official EPA AirData files. EPA source facts remain public domain. The authored tabular compilation and documentation are CC BY 4.0; Python code and the bundled wheel are Apache-2.0. See release/RIGHTS_AND_SOURCES.md.

SenseShift-RL is independent and is not endorsed, approved, or certified by EPA. Its sensing costs and rewards are benchmark definitions, not real operational quantities.

Companion work

• Kaggle dataset
• Quick Start and World Audit
• Baselines and Shift Evaluation