cognitive/Things/Generic_POMDP/configuration.yaml

# Generic POMDP Configuration

# Model Parameters
model:
  name: "Generic POMDP"
  description: "Generic POMDP implementation using Active Inference"
  version: "1.0.3"
  temporal_preferences: true  # Enable temporal preference handling
  learning_mode: "continuous"  # Enable continuous learning

# Space Dimensions
dimensions:
  observations: 5
  states: 4
  actions: 3
  total_timesteps: 10  # Total number of timesteps to run simulation
  planning_horizon: 4   # Number of timesteps to look ahead for policy evaluation
  memory_length: 5     # Number of past observations to consider for belief updates
  temporal_discount: 0.95  # Discount factor for future preferences
  belief_momentum: 0.8  # Added belief momentum for smoother updates

# Matrix Specifications
matrices:
  A_matrix:  # Observation model P(o|s)
    initialization: "random_stochastic"
    constraints:
      - "column_stochastic"
      - "non_negative"
      - "sparse"  # Added sparsity constraint for more distinct state-observation mappings
    sparsity_params:
      target_sparsity: 0.7
      min_value: 0.05
    validation:
      max_condition_number: 1e6
      min_eigenvalue: 1e-6
    learning:
      enabled: true
      rate: 0.1
      momentum: 0.9
      
  B_matrix:  # Transition model P(s'|s,a)
    initialization: "identity_based"
    initialization_params:
      strength: 0.8  # Strength of identity component
      noise: 0.1    # Added noise parameter for exploration
      temporal_coherence: 0.9  # Higher values make state transitions more temporally coherent
    constraints:
      - "column_stochastic"
      - "non_negative"
      - "temporal_consistency"  # Added constraint for temporal consistency
    validation:
      max_condition_number: 1e6
      min_eigenvalue: 1e-6
    learning:
      enabled: true
      rate: 0.05
      momentum: 0.95
      
  C_matrix:  # Preferences over observations
    initialization: "temporal_goal_directed"  # Changed to temporal goal-directed initialization
    initialization_params:
      default_value: 0.0
      goal_states: [0, 3]  # Specify preferred goal states
      goal_value: 1.0
      avoid_value: -1.0
      temporal_profile: "increasing"  # Preferences increase over time
      temporal_scale: 1.5  # Scale factor for temporal preference increase
    constraints:
      - "finite_values"
      - "bounded"
      - "temporal_coherence"  # Added temporal coherence constraint
    bounds:
      min: -2.0
      max: 2.0
    temporal_params:
      horizon_weight: 1.2  # Weight future preferences more heavily
      smoothing: 0.1  # Smooth preference transitions
    learning:
      enabled: true
      rate: 0.2
      momentum: 0.8
      
  D_matrix:  # Prior beliefs over states
    initialization: "informed"  # Changed to informed initialization
    initialization_params:
      concentration: 1.0  # Dirichlet concentration parameter
      temporal_bias: 0.2  # Bias towards temporally consistent states
    constraints:
      - "normalized"
      - "non_negative"
      - "minimum_entropy"  # Added minimum entropy constraint
    min_entropy: 0.5
    learning:
      enabled: true
      rate: 0.15
      momentum: 0.85
      
  E_matrix:  # Prior over policies
    initialization: "temporal_softmax"  # Changed to temporal-aware softmax
    initialization_params:
      temperature: 1.0
      temporal_bonus: 0.2  # Bonus for temporally coherent policies
    constraints:
      - "normalized"
      - "non_negative"
      - "entropy_regularized"  # Added entropy regularization
    entropy_weight: 0.1
    learning:
      enabled: true
      rate: 0.1
      momentum: 0.9

# Inference Parameters
inference:
  learning_rate: 0.05  # Reduced learning rate for more stable updates
  temperature: 0.8    # Reduced temperature for more focused exploration
  convergence_threshold: 1e-6
  max_iterations: 100
  belief_update:
    method: "variational"  # Changed to variational inference
    momentum: 0.9
    adaptive_lr: true
    min_lr: 1e-4
    max_lr: 0.5
    window_size: 10
    smoothing_factor: 0.2  # Added smoothing for belief updates
    regularization: 0.01   # Added regularization
  policy_selection:
    method: "temporal_softmax"
    temperature: 1.0
    exploration_bonus: 0.1
    temporal_horizon_bonus: 0.2  # Bonus for policies with good long-term outcomes

# Learning Parameters
learning:
  enabled: true
  type: "online"
  parameters:
    memory_decay: 0.95
    learning_rate_decay: 0.999
    min_learning_rate: 1e-4
    exploration_decay: 0.995
    min_exploration: 0.05
    belief_momentum: 0.9     # Added belief momentum
    temporal_smoothing: 0.2  # Added temporal smoothing
  regularization:
    type: "l2"
    strength: 0.01
    temporal_coherence: 0.1  # Added temporal coherence regularization
  curriculum:
    enabled: true
    difficulty_increase_rate: 0.1
    max_difficulty: 1.0
    adaptive_pacing: true    # Added adaptive pacing
  temporal:
    sequence_length: 5
    prediction_horizon: 3
    sequence_weight: 0.3
    consistency_weight: 0.2  # Added consistency weight
  belief_update:
    method: "momentum"       # Changed to momentum-based updates
    momentum: 0.9
    step_size: 0.1
    regularization: 0.01

# Analysis Parameters
analysis:
  enabled: true
  metrics:
    - "belief_entropy"
    - "free_energy"
    - "accuracy"
    - "temporal_consistency"
    - "preference_satisfaction"
    - "learning_progress"    # Added learning progress metric
  temporal_analysis:
    window_size: 5
    overlap: 2
    metrics:
      - "state_transitions"
      - "observation_sequences"
      - "belief_trajectories"
      - "learning_curves"    # Added learning curves
  information_theory:
    compute_mutual_info: true
    compute_kl_divergence: true
    compute_entropy_rate: true
    temporal_dependencies: true  # Added temporal dependencies

# Numerical Parameters
numerical:
  stability_threshold: 1e-12
  max_condition_number: 1e6
  gradient_clip: 10.0
  belief_clip: [1e-7, 1.0]
  precision_scaling: true
  precision_params:
    initial: 1.0
    learning_rate: 0.1
    min_value: 0.1
    max_value: 10.0

# Output Settings
output:
  base_dir: "Output/"
  subdirs:
    logs: "logs/"
    plots: "plots/"
    test_results: "test_results/"
    simulations: "simulations/"
    checkpoints: "checkpoints/"  # Added checkpoints directory
    analysis: "analysis/"  # Added analysis directory
  file_formats:
    plots: ["png", "pdf"]
    data: ["csv", "json"]
    checkpoints: ["pt", "npz"]
    analysis: ["json", "yaml"]
  logging:
    level: "INFO"
    format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
    file: "Output/logs/simulation.log"
    rotation: "1 day"
    backup_count: 7

# Visualization Settings
visualization:
  enabled: true
  output_dir: "Output/plots/"
  formats:
    - "png"
    - "pdf"
  required_plots:
    - "belief_evolution"
    - "free_energy"
    - "action_probabilities"
    - "observation_counts"
    - "learning_curves"      # Added learning curves
    - "belief_trajectories"  # Added belief trajectories
    - "state_transitions"    # Added state transition visualization
    - "temporal_preference_heatmap"  # Added temporal preference visualization
    - "policy_evaluation_over_time"  # Added policy evaluation visualization
    - "information_theory_metrics"   # Added information theory metrics
    - "temporal_consistency_analysis" # Added temporal consistency analysis
  style:
    colormap: "viridis"
    figure_size: [10, 6]
    font_size: 12
    dpi: 300
    grid: true
    legend_location: "best"
  interactive:
    enabled: true
    backend: "plotly"
  temporal_plots:
    enabled: true
    types:
      - "state_sequence_diagram"
      - "belief_flow"
      - "preference_evolution"
      - "policy_tree"
    animation:
      enabled: true
      fps: 2
      duration: 10

# Testing Configuration
testing:
  enabled: true
  coverage_threshold: 90
  test_cases:
    initialization:
      - "matrix_properties"
      - "state_initialization"
      - "constraint_satisfaction"
      - "temporal_consistency"  # Added temporal consistency test
    dynamics:
      - "belief_updates"
      - "action_selection"
      - "observation_generation"
      - "temporal_consistency"
      - "sequence_prediction"  # Added sequence prediction test
    learning:
      - "belief_convergence"
      - "policy_improvement"
      - "exploration_decay"
      - "temporal_credit_assignment"  # Added temporal credit assignment test
    convergence:
      - "free_energy_minimization"
      - "belief_convergence"
      - "learning_stability"
      - "temporal_stability"  # Added temporal stability test
    numerical:
      - "stability"
      - "precision"
      - "gradient_properties"
      - "temporal_coherence"  # Added temporal coherence test
  benchmarks:
    enabled: true
    metrics:
      - "belief_update_time"
      - "policy_evaluation_time"
      - "learning_convergence_rate"
      - "temporal_prediction_accuracy"  # Added temporal prediction metric
    baseline_performance:
      max_update_time_ms: 50
      min_convergence_rate: 0.1
      min_temporal_accuracy: 0.7  # Added temporal accuracy threshold