# =============================================================================
# Experiment config: enhanced ML/DL tuning + automatic HTML report
# =============================================================================

# 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
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]

# Experiment dimensions
source_models: [RegCM, clWRF]
leads: [1, 2, 3, 4, 5, 6]
targets: [pr_o, t2_o]
spatial_modes: [regional, national]

# Keep P2 and P4 as the main predictor sets because the previous experiment showed
# that P_fortran22 was not consistently better. Add P_fortran22 here if needed.
predictor_sets: [P2, P4]
# predictor_sets: [P2, P4, P_fortran22]

# Ridge is the benchmark. XGBoost and MLP/ANN are tested with several variants.
# CNN is optional and disabled by default because it requires PyTorch.
methods:
  - raw
  - ridge
  - xgb_d2
  - xgb_d3
  - xgb_d4
  - xgb_slow
  - xgb_regularized
  - mlp_small
  - mlp_medium
  - mlp_deep
  - ann_tanh
  - ann_lbfgs
# Optional, only if PyTorch is installed and tested:
#  - cnn1d_small

# Split by target year
train_years: [1982, 2013]
test_years: [2014, 2021]

# Modeling style
# monthly_anomaly trains separate models for each calendar month and uses anomalies
# relative to training-period monthly climatology.
modeling_style: monthly_anomaly
include_lat_lon: true
include_region_feature: true
clip_precipitation_to_zero: true
rmae_percent: true

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

# Ridge benchmark
ridge_alpha: 1.0

# XGBoost global parameters and variants. Each task uses n_jobs=1 to avoid
# oversubscription because SLURM runs many tasks in parallel.
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/ANN base parameters and variants. The lbfgs variant is often useful for
# small tabular datasets and can be more stable than Adam in short samples.
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

# Optional CNN settings. Requires PyTorch; not installed by requirements.txt.
cnn1d_small_params:
  n_epochs: 120
  batch_size: 512
  learning_rate: 0.001
  weight_decay: 0.0001
  channels: 16
  patience: 20
  validation_fraction: 0.15