# =============================================================================
# PA3 experiment: ML/DL bias correction + year-blocked stacking ensemble
# =============================================================================

# Paths
data_dir: /home/tanpv/Progs/562_VNU/Data/For_BC
mask_file: /home/tanpv/Data/AWC/Mask_VN_0.25_Corrected.nc
output_dir: outputs
file_pattern: "lead{lead}_{source_model}_01_BL_Test_VN.nc"

# Data conventions
year_start: 1982
year_end: 2021
missing_value: -99.0
valid_regions: [1, 2, 3, 4, 5, 6, 7]

# Main experiment dimensions
source_models: [RegCM, clWRF]
leads: [1, 2, 3, 4, 5, 6]
targets: [pr_o, t2_o]
spatial_modes: [regional, national]
predictor_sets: [P2, P4]

# Individual base-correction methods. Ridge is the benchmark.
methods:
  - raw
  - ridge
  - xgb_d2
  - xgb_d3
  - xgb_d4
  - xgb_slow
  - xgb_regularized
  - mlp_small
  - mlp_medium
  - mlp_deep
  - ann_tanh
  - ann_lbfgs

# Split by target year. Metrics are computed only on the independent test period.
train_years: [1982, 2013]
test_years: [2014, 2021]

# Bias-correction formulation
modeling_style: monthly_anomaly
include_lat_lon: true
include_region_feature: true
clip_precipitation_to_zero: true

# Runtime
n_jobs: 20
overwrite: false
save_predictions: true
random_seed: 42

# =============================================================================
# Metrics rule
# =============================================================================
# pr_o: RMAE_percent is the primary absolute-error metric; CORR is secondary.
# t2_o: MAE (deg C) is the primary absolute-error metric; RMSE/CORR are secondary.
rmae_percent: true
primary_metrics:
  pr_o: [RMAE_percent, CORR]
  t2_o: [MAE, RMSE, CORR]

# =============================================================================
# Individual model hyperparameters
# =============================================================================
ridge_alpha: 1.0

xgboost_base_params:
  objective: reg:squarederror
  n_jobs: 1
  tree_method: hist
  verbosity: 0

xgboost_variants:
  xgb_d2:
    n_estimators: 500
    max_depth: 2
    learning_rate: 0.03
    subsample: 0.90
    colsample_bytree: 0.90
    reg_lambda: 5.0
    reg_alpha: 0.0
    min_child_weight: 3.0
  xgb_d3:
    n_estimators: 500
    max_depth: 3
    learning_rate: 0.03
    subsample: 0.90
    colsample_bytree: 0.90
    reg_lambda: 5.0
    reg_alpha: 0.0
    min_child_weight: 3.0
  xgb_d4:
    n_estimators: 500
    max_depth: 4
    learning_rate: 0.03
    subsample: 0.85
    colsample_bytree: 0.85
    reg_lambda: 8.0
    reg_alpha: 0.0
    min_child_weight: 4.0
  xgb_slow:
    n_estimators: 900
    max_depth: 3
    learning_rate: 0.015
    subsample: 0.90
    colsample_bytree: 0.90
    reg_lambda: 8.0
    reg_alpha: 0.0
    min_child_weight: 3.0
  xgb_regularized:
    n_estimators: 700
    max_depth: 2
    learning_rate: 0.02
    subsample: 0.75
    colsample_bytree: 0.75
    reg_lambda: 15.0
    reg_alpha: 0.5
    min_child_weight: 5.0

mlp_base_params:
  max_iter: 700
  early_stopping: true
  validation_fraction: 0.15
  n_iter_no_change: 40
  learning_rate_init: 0.001
  alpha: 0.001
  solver: adam
  activation: relu

mlp_variants:
  mlp_small:
    hidden_layer_sizes: [32]
    alpha: 0.001
  mlp_medium:
    hidden_layer_sizes: [64, 32]
    alpha: 0.001
  mlp_deep:
    hidden_layer_sizes: [128, 64, 32]
    alpha: 0.003
  ann_tanh:
    hidden_layer_sizes: [64, 32]
    activation: tanh
    alpha: 0.001
  ann_lbfgs:
    hidden_layer_sizes: [32]
    activation: tanh
    solver: lbfgs
    alpha: 0.01
    max_iter: 700

# =============================================================================
# Stacking ensemble options
# =============================================================================
stacking:
  enabled: true
  # Use year-blocked CV inside training years to generate out-of-fold base predictions.
  # This avoids training the meta-model on predictions from models that already saw those samples.
  cv_type: year_kfold
  n_splits: 8
  shuffle_years: false

  # Build three products:
  #   within_RegCM: stack only RegCM base models
  #   within_clWRF: stack only clWRF base models
  #   multi_source: stack RegCM + clWRF base models together
  scopes: [within_RegCM, within_clWRF, multi_source]

  # Base models used by stacking. Keep this list shorter than the full methods list to limit cost.
  base_methods: [ridge, xgb_d2, xgb_regularized, mlp_small, ann_tanh]
  base_predictor_sets: [P2, P4]

  # Meta-models to test. stack_mean has no fitted parameters; stack_nnls uses non-negative linear weights.
  meta_methods: [stack_mean, stack_ridge, stack_elasticnet, stack_nnls]
  stack_ridge_alpha: 1.0
  stack_elasticnet_alpha: 0.001
  stack_elasticnet_l1_ratio: 0.25

  # Optional geographical features at meta level.
  include_meta_geo_features: false

  # Minimum number of samples required to train a meta-model for a region/month.
  min_meta_samples: 30