juksentang/BIRL-Climate-Adaptation

Bayesian Inverse Reinforcement Learning for Smallholder Climate Adaptation — 6 Sub-Saharan African Countries ## Acknowledgments Cloud computing resources were provided by the Google Cloud TPU Research Cloud (TRC) program. Geospatial data extraction was supported by the Google Earth Engine (GEE) academic research quota. # BIRL Formal Analysis Pipeline Bayesian Inverse Reinforcement Learning for Smallholder Agricultural Decision-Making. 6 Sub-Saharan African countries (Ethiopia, Malawi, Mali, Nigeria, Tanzania, Uganda), 2008–2023. ## Directory Structure ``` Formal Analysis/ ├── README.md │ ├── data/ ← Shared data │ ├── all_countries_panel_birl.parquet ← Full panel (514K × 211) │ └── birl_sample.parquet ← Analysis sample (222K × 244) │ ├── 01_Data_Screening/ ← Sample selection & cleaning ├── 02_Action_Space/ ← 27 actions (9 crops × 3 intensity) ├── 03_FDH/ ← Order-m FDH frontier estimation ├── 04_Env_Model/ ← LightGBM environment model (Colab) ├── 05_BIRL_SVI/ ← Variational Inference prototype (Colab) ├── 06_BIRL_MCMC/ ← MCMC posterior inference (GCP) ├── 07_2050_Counter_Fact/ ← 2050 climate counterfactual & policy welfare │ └── docs/ ← Build guides, reports, and analysis notes ``` ## Pipeline Overview ``` all_countries_panel_birl.parquet (514K × 211) │ ▼ Steps 01-03 (local, ~30s) birl_sample.parquet (222K × 244, with actions + FDH efficiency) │ ▼ Step 04 (Colab, ~2.5h) env_model_output.npz + model_mu.txt + model_sigma.txt │ ▼ Step 05→06 (Colab/GCP, hours, Complete analysis requires 8Chips TPU V4) posterior.pkl (MCMC: ρ_c, γ_c per country, 12K samples) │ ▼ Step 07 (local, ~30min) CE tables, climate loss, policy value, synergy ``` ## Pipeline Steps | Step | Directory | Runner | Environment | Time | |------|-----------|--------|-------------|-----:| | 01 | `01_Data_Screening/` | `screen_and_clean.py` | Local | ~6s | | 02 | `02_Action_Space/` | `build_ …