# Hyperparameter Tuning Configuration for Foxhunt ML Models - Agent 49 Optimized
# Defines focused search spaces based on Agent 49 specifications

# Global tuning settings
global:
  optimization_direction: maximize  # maximize sharpe_ratio
  pruning_enabled: true
  median_pruner:
    n_startup_trials: 5      # No pruning for first 5 trials (establish baseline)
    n_warmup_steps: 10       # Wait 10 epochs before starting to prune
    interval_steps: 5        # Check for pruning every 5 epochs
  sampler: TPE  # Tree-structured Parzen Estimator (supports Bayesian optimization)

# Model-specific search spaces (Agent 49 Specifications)
models:
  # DQN: Deep Q-Network for Reinforcement Learning
  DQN:
    # Agent 49 specified parameters
    learning_rate:
      type: categorical
      choices: [0.00001, 0.0001, 0.001]  # [1e-5, 1e-4, 1e-3]
    batch_size:
      type: categorical
      choices: [64, 128, 256]
    gamma:
      type: categorical
      choices: [0.95, 0.99, 0.999]

    # Additional DQN-specific parameters (for completeness)
    epochs:
      type: int
      low: 100
      high: 300
      step: 50
    replay_buffer_size:
      type: categorical
      choices: [50000, 100000]
    epsilon_start:
      type: float
      low: 0.95
      high: 1.0
      step: 0.05
    epsilon_end:
      type: float
      low: 0.01
      high: 0.05
      step: 0.01
    epsilon_decay_steps:
      type: int
      low: 5000
      high: 10000
      step: 1000
    target_update_frequency:
      type: categorical
      choices: [500, 1000]
    use_prioritized_replay:
      type: categorical
      choices: [true, false]

  # PPO: Proximal Policy Optimization
  PPO:
    # Agent 49 specified parameters
    learning_rate:
      type: categorical
      choices: [0.00003, 0.0001, 0.0003]  # [3e-5, 1e-4, 3e-4]
    entropy_coef:
      type: categorical
      choices: [0.01, 0.05, 0.1]
    clip_ratio:
      type: categorical
      choices: [0.1, 0.2, 0.3]

    # Additional PPO-specific parameters (for completeness)
    epochs:
      type: int
      low: 100
      high: 300
      step: 50
    batch_size:
      type: categorical
      choices: [128, 256, 512]
    value_loss_coef:
      type: categorical
      choices: [0.5, 1.0]
    rollout_steps:
      type: categorical
      choices: [512, 1024, 2048]
    minibatch_size:
      type: categorical
      choices: [64, 128, 256]
    gae_lambda:
      type: categorical
      choices: [0.95, 0.97, 0.99]

  # MAMBA-2: State Space Model
  MAMBA_2:
    # Agent 49 specified parameters
    learning_rate:
      type: categorical
      choices: [0.00001, 0.0001, 0.001]  # [1e-5, 1e-4, 1e-3]
    state_dim:
      type: categorical
      choices: [16, 32, 64]  # State size in Agent 49 spec
    num_layers:
      type: categorical
      choices: [4, 6, 8]  # Layers in Agent 49 spec

    # Additional MAMBA-2 parameters (for completeness)
    epochs:
      type: int
      low: 50
      high: 150
      step: 25
    batch_size:
      type: categorical
      choices: [64, 128, 256]
    hidden_dim:
      type: categorical
      choices: [128, 256, 512]
    dt_min:
      type: float
      low: 0.0001
      high: 0.001
      log: true
    dt_max:
      type: float
      low: 0.01
      high: 0.1
      log: true
    use_cuda_kernels:
      type: categorical
      choices: [true, false]

  # TFT: Temporal Fusion Transformer
  TFT:
    # Agent 49 specified parameters
    learning_rate:
      type: categorical
      choices: [0.00001, 0.0001, 0.001]  # [1e-5, 1e-4, 1e-3]
    num_heads:
      type: categorical
      choices: [4, 8, 16]  # Attention heads in Agent 49 spec
    hidden_dim:
      type: categorical
      choices: [128, 256, 512]  # Hidden dim in Agent 49 spec

    # Additional TFT parameters (for completeness)
    epochs:
      type: int
      low: 50
      high: 150
      step: 25
    batch_size:
      type: categorical
      choices: [64, 128, 256]
    num_layers:
      type: categorical
      choices: [3, 4, 6]
    lookback_window:
      type: categorical
      choices: [30, 50, 100]
    forecast_horizon:
      type: categorical
      choices: [5, 10, 20]
    dropout_rate:
      type: categorical
      choices: [0.1, 0.2, 0.3]

# Search space sizes (for reporting):
# DQN: 3 (lr) * 3 (batch) * 3 (gamma) = 27 combinations (grid) + continuous params
# PPO: 3 (lr) * 3 (entropy) * 3 (clip) = 27 combinations (grid) + continuous params
# MAMBA-2: 3 (lr) * 3 (state) * 3 (layers) = 27 combinations (grid) + continuous params
# TFT: 3 (lr) * 3 (heads) * 3 (hidden) = 27 combinations (grid) + continuous params
#
# Total grid combinations per model: 27
# Recommended trials per model: 50-100 (TPE sampler explores beyond grid)