%% @doc Meta-Controller record definitions for LTC-based hyperparameter optimization
%%
%% This module implements a meta-learning system that uses Liquid Time-Constant
%% (LTC) networks to dynamically control neuroevolution hyperparameters.
%%
%% The meta-controller operates at a higher timescale than the task networks,
%% learning to "optimize the optimizer" by adapting parameters based on
%% training dynamics.
%%
%% @author Macula.io
%% @copyright 2025 Macula.io

-ifndef(META_CONTROLLER_HRL).
-define(META_CONTROLLER_HRL, true).

%%% ============================================================================
%%% Types
%%% ============================================================================

-type meta_param() :: mutation_rate | mutation_strength | selection_ratio.
-type reward_component() :: convergence_speed | final_fitness | efficiency_ratio |
                           diversity_aware | normative_structure.

%%% ============================================================================
%%% Configuration Record
%%% ============================================================================

%% @doc Configuration for the meta-controller.
%%
%% Controls the LTC network architecture and reward signal composition.
-record(meta_config, {
    %% Network topology: {Inputs, HiddenLayers, Outputs}
    %% Default: {8, [16, 8], 4} - 8 inputs (training metrics), 4 outputs (params)
    network_topology = {8, [16, 8], 4} :: {pos_integer(), [pos_integer()], pos_integer()},

    %% LTC neuron type: cfc (fast) or ltc (accurate ODE)
    neuron_type = cfc :: ltc | cfc,

    %% Base time constant for meta-controller neurons
    %% Higher values = slower adaptation = more stable
    time_constant = 50.0 :: float(),

    %% State bound for LTC neurons
    state_bound = 1.0 :: float(),

    %% Reward component weights (must sum to ~1.0)
    reward_weights = #{
        convergence_speed => 0.25,
        final_fitness => 0.25,
        efficiency_ratio => 0.20,
        diversity_aware => 0.15,
        normative_structure => 0.15
    } :: #{reward_component() => float()},

    %% Learning rate for gradient-based meta-training
    learning_rate = 0.001 :: float(),

    %% Parameter bounds: {ParamName, {Min, Max}}
    param_bounds = #{
        mutation_rate => {0.01, 0.5},
        mutation_strength => {0.05, 1.0},
        selection_ratio => {0.10, 0.50},
        evaluations_per_individual => {3, 20}
    } :: #{meta_param() => {float(), float()}},

    %% Whether to include population_size as a controllable parameter
    %% (disabled by default as it affects computation more than other params)
    control_population_size = false :: boolean(),

    %% History window size for computing reward signals
    history_window = 10 :: pos_integer(),

    %% Momentum for parameter updates (smooths changes)
    momentum = 0.9 :: float()
}).

-type meta_config() :: #meta_config{}.

%%% ============================================================================
%%% State Record
%%% ============================================================================

%% @doc Internal state for the meta-controller gen_server.
-record(meta_state, {
    %% Server identifier
    id :: term(),

    %% Configuration
    config :: meta_config(),

    %% Current hyperparameter values
    current_params = #{
        mutation_rate => 0.10,
        mutation_strength => 0.30,
        selection_ratio => 0.20
    } :: #{meta_param() => float()},

    %% LTC network internal states (one per hidden neuron)
    %% Maps neuron_id => internal_state
    ltc_states = #{} :: #{term() => float()},

    %% LTC network weights
    %% Structure: #{layer => #{neuron_id => [{input_id, weight}]}}
    ltc_weights = #{} :: map(),

    %% Training metrics history (circular buffer of last N generations)
    metrics_history = [] :: [generation_metrics()],

    %% Parameter change momentum (for smooth updates)
    param_momentum = #{} :: #{meta_param() => float()},

    %% Generation counter
    generation = 0 :: non_neg_integer(),

    %% Cumulative reward for current training run
    cumulative_reward = 0.0 :: float(),

    %% Best fitness ever seen (for normalization)
    best_fitness_ever = 0.0 :: float(),

    %% Stagnation counter (generations without improvement)
    stagnation_count = 0 :: non_neg_integer(),

    %% Running state
    running = false :: boolean()
}).

-type meta_state() :: #meta_state{}.

%%% ============================================================================
%%% Generation Metrics Record
%%% ============================================================================

%% @doc Metrics captured from a single generation for meta-learning.
-record(generation_metrics, {
    %% Generation number
    generation :: pos_integer(),

    %% Fitness metrics
    best_fitness :: float(),
    avg_fitness :: float(),
    worst_fitness :: float(),
    fitness_std_dev :: float(),

    %% Improvement metrics
    fitness_delta :: float(),          %% Change from previous generation
    relative_improvement :: float(),   %% Percentage improvement

    %% Diversity metrics
    population_diversity :: float(),   %% Variance in fitness
    strategy_entropy :: float(),       %% Information content of strategies

    %% Efficiency metrics
    evaluations_used :: pos_integer(),
    fitness_per_evaluation :: float(),

    %% Normative structure metrics
    diversity_corridors :: float(),     %% Cluster distinctness
    adaptation_readiness :: float(),    %% Variance in adjacent traits

    %% Current parameters used
    params_used :: #{meta_param() => float()},

    %% Timestamp
    timestamp :: erlang:timestamp()
}).

-type generation_metrics() :: #generation_metrics{}.

%%% ============================================================================
%%% Reward Signal Record
%%% ============================================================================

%% @doc Computed reward signal for meta-controller training.
-record(meta_reward, {
    %% Individual reward components
    convergence_speed :: float(),
    final_fitness :: float(),
    efficiency_ratio :: float(),
    diversity_aware :: float(),
    normative_structure :: float(),

    %% Composite reward (weighted sum)
    total :: float(),

    %% Generation this reward is for
    generation :: pos_integer()
}).

-type meta_reward() :: #meta_reward{}.

%%% ============================================================================
%%% Network Structure Records
%%% ============================================================================

%% @doc LTC neuron state for the meta-controller network.
-record(meta_neuron, {
    %% Neuron identifier: {layer, index}
    id :: {pos_integer(), pos_integer()},

    %% Internal state x(t)
    internal_state = 0.0 :: float(),

    %% Base time constant
    time_constant :: float(),

    %% State bound
    state_bound :: float(),

    %% Input weights: [{source_id, weight}]
    input_weights = [] :: [{term(), float()}],

    %% Bias
    bias = 0.0 :: float(),

    %% Backbone weights for CfC f() function
    backbone_weights = [] :: [float()],

    %% Head weights for CfC h() function
    head_weights = [] :: [float()]
}).

-type meta_neuron() :: #meta_neuron{}.

%% @doc Output mapping for meta-controller.
%% Maps network outputs to parameter adjustments.
-record(output_mapping, {
    %% Output index -> parameter name
    index_to_param :: #{non_neg_integer() => meta_param()},

    %% Parameter name -> output index
    param_to_index :: #{meta_param() => non_neg_integer()},

    %% Activation functions for each output (e.g., sigmoid for bounded params)
    output_activations :: #{non_neg_integer() => atom()}
}).

-type output_mapping() :: #output_mapping{}.

%%% ============================================================================
%%% Training Event Record
%%% ============================================================================

%% @doc Training event for gradient-based meta-learning.
-record(meta_training_event, {
    %% Generation
    generation :: pos_integer(),

    %% Input features (normalized metrics)
    inputs :: [float()],

    %% Output parameters (raw network outputs before scaling)
    outputs :: [float()],

    %% Reward received
    reward :: float(),

    %% Gradient estimate for this step
    gradients :: #{term() => float()}
}).

-type meta_training_event() :: #meta_training_event{}.

%%% ============================================================================
%%% Structure Metrics Record (Normative Awareness)
%%% ============================================================================

%% @doc Structure metrics for normative awareness reward component.
%%
%% These metrics capture whether the population maintains the capacity
%% for future adaptation, not just current fitness.
-record(structure_metrics, {
    %% Number of distinct strategy clusters in population
    diversity_corridors = 0.0 :: float(),

    %% Variance in traits that are adjacent to fitness (exploration potential)
    adaptation_readiness = 0.0 :: float(),

    %% Estimated distance to unexplored fitness regions
    breakthrough_potential = 0.0 :: float(),

    %% Shannon entropy of strategy distribution
    strategy_entropy = 0.0 :: float()
}).

-type structure_metrics() :: #structure_metrics{}.

-endif. %% META_CONTROLLER_HRL