src/engine/selection.py

import numpy as np
from numba import njit, prange
from src.config import cfg

# Costanti di dominio (mapping per la vettorializzazione)
CODE_OFF = -1
CODE_ABSENCE = -2
CONS_TYPE_HARD = 1
CONS_TYPE_SOFT = 2
CONS_TYPE_ABSENCE = 3

@njit(cache=True)
def get_valid_start_slots_numba(day_idx, shift_len, cons_type, cons_val, closing_slot, is_closed):
    """
    Calcolo degli slot di inizio turno validi. 
    JIT-compiled per massimizzare il throughput. Ritorna strictly un ndarray.
    """
    # 1. Vincoli Hard (Assenze e Turni Fissi)
    if cons_type == CONS_TYPE_ABSENCE:
        return np.array([CODE_ABSENCE], dtype=np.int64)
        
    if cons_type == CONS_TYPE_HARD:
        target_slot = cons_val
        # Boundary check per evitare sforamenti oltre l'orario di chiusura
        if is_closed or (target_slot + shift_len > closing_slot):
            return np.array([CODE_OFF], dtype=np.int64)
        return np.array([target_slot], dtype=np.int64)

    # 2. Assegnazione Standard
    if is_closed:
        return np.empty(0, dtype=np.int64) 

    max_start = closing_slot - shift_len
    if max_start < 0:
        return np.empty(0, dtype=np.int64)
        
    # Generazione array di slot contigui validi
    return np.arange(0, max_start + 1, dtype=np.int64)

@njit(cache=True)
def calculate_fitness_numba(individual, masks, lengths, target_days_arr, cons_types, cons_vals, target_demand, weights, daily_slots):
    """
    Motore core di valutazione della fitness. 
    Aggrega i vincoli operativi (contratti, weekend, preferenze) e vettorializza le penalità.
    """
    num_emps = len(lengths)
    num_days = 7
    num_slots = daily_slots
    
    current_coverage = np.zeros((num_days, num_slots), dtype=np.int64)
    
    homogeneity_penalty = 0.0
    soft_pref_penalty = 0.0
    contract_penalty = 0.0
    equity_penalty = 0.0
    
    W_UNDER = weights[0]
    W_OVER  = weights[1]
    W_HOMO  = weights[2]
    W_SOFT  = weights[3]
    
    PENALTY_PER_DAY_MISMATCH = W_UNDER * 2.0 

    sum_wk_work = 0.0
    sum_sq_wk_work = 0.0

    # Scansione matrice turni per estrazione metriche operative
    for i in range(num_emps):
        mask_len = lengths[i]
        real_mask = masks[i, :mask_len]
        
        days_worked_count = 0 
        weekend_days = 0
        
        sum_slots = 0.0
        sum_sq_slots = 0.0
        count_slots = 0

        for day in range(num_days):
            start = individual[i, day]
            
            if start >= 0:
                days_worked_count += 1
                sum_slots += start
                sum_sq_slots += start**2
                count_slots += 1
                if day >= 5: weekend_days += 1
                
                end = min(start + mask_len, num_slots)
                real_len = end - start
                
                # Proiezione del fenotipo (maschera VDT) sulla coverage matrix
                if real_len > 0:
                    current_coverage[day, start:end] += real_mask[:real_len]
            
            # Valutazione desiderata (Soft Constraint)
            if cons_types[i, day] == CONS_TYPE_SOFT and start >= 0:
                target_slot = cons_vals[i, day]
                if start != target_slot:
                    soft_pref_penalty += abs(start - target_slot)

        # Check target contrattuale (mix lavorativi vs riposi)
        tgt = target_days_arr[i]
        if days_worked_count != tgt:
            contract_penalty += (abs(days_worked_count - tgt) * PENALTY_PER_DAY_MISMATCH)

        # Calcolo varianza per penalizzare pattern orari troppo frammentati
        if count_slots > 1:
            mean = sum_slots / count_slots
            var = (sum_sq_slots / count_slots) - (mean**2)
            if var > 0: homogeneity_penalty += np.sqrt(var)

        sum_wk_work += weekend_days
        sum_sq_wk_work += weekend_days**2

    # Equità aziendale: bilanciamento dei carichi sui weekend tra dipendenti
    if num_emps > 1:
        mean_wk = sum_wk_work / num_emps
        var_wk = (sum_sq_wk_work / num_emps) - (mean_wk**2)
        if var_wk > 0:
            equity_penalty = np.sqrt(var_wk) * W_HOMO * 10.0

    # Calcolo delta rispetto al fabbisogno (Demand vs Coverage)
    diff = current_coverage - target_demand
    
    under_score = 0.0
    flattened_diff = diff.flatten()
    for val in flattened_diff:
        if val < 0: under_score += abs(val)
    under_score *= W_UNDER
    
    over_score = 0.0
    safe_target = target_demand.flatten()
    for k in range(len(flattened_diff)):
        val = flattened_diff[k]
        if val > 0:
            tgt_val = safe_target[k]
            if tgt_val == 0: tgt_val = 1
            over_score += (val / tgt_val)
    over_score *= (W_OVER * 10.0)

    # Aggregazione Loss Function
    total_score = (under_score + 
                   over_score + 
                   (homogeneity_penalty * W_HOMO) +
                   (soft_pref_penalty * W_SOFT) +
                   contract_penalty + 
                   equity_penalty)
    
    return total_score, current_coverage

@njit(parallel=True, cache=True) 
def calculate_population_fitness_parallel(population, masks, lengths, target_days_arr, cons_types, cons_vals, target_demand, weights, daily_slots):
    """
    Valutazione massiva della popolazione. Sfrutta il multithreading (prange) di Numba.
    """
    pop_size = len(population)
    scores = np.zeros(pop_size, dtype=np.float64)
    for i in prange(pop_size):
        sc, _ = calculate_fitness_numba(
            population[i], 
            masks, lengths, target_days_arr, cons_types, cons_vals, 
            target_demand, weights, daily_slots
        )
        scores[i] = sc
    return scores

def tournament_selection(population, fitness_scores):
    """Selezione a torneo standard."""
    k = cfg.genetic_params.get('tournament_size', 5)
    indices = np.random.choice(len(population), k, replace=False)
    best_idx = indices[np.argmin(fitness_scores[indices])]
    return population[best_idx]

@njit(cache=True)
def calculate_detailed_score(individual, masks, lengths, target_days_arr, cons_types, cons_vals, target_demand, weights, daily_slots):
    """
    Versione verbosa del calcolo fitness usata post-convergenza 
    per l'estrazione delle metriche finali di business da mostrare in UI.
    """
    num_emps = len(lengths)
    num_days = 7
    num_slots = daily_slots
    
    current_coverage = np.zeros((num_days, num_slots), dtype=np.int64)
    
    homogeneity_penalty = 0.0
    soft_pref_penalty = 0.0
    contract_penalty = 0.0
    equity_penalty = 0.0
    
    W_UNDER = weights[0]
    W_OVER  = weights[1]
    W_HOMO  = weights[2]
    W_SOFT  = weights[3]
    
    PENALTY_PER_DAY_MISMATCH = W_UNDER * 2.0 

    sum_wk_work = 0.0
    sum_sq_wk_work = 0.0

    for i in range(num_emps):
        mask_len = lengths[i]
        real_mask = masks[i, :mask_len]
        days_worked_count = 0 
        weekend_days = 0
        sum_slots = 0.0
        sum_sq_slots = 0.0
        count_slots = 0

        for day in range(num_days):
            start = individual[i, day]
            if start >= 0:
                days_worked_count += 1
                sum_slots += start
                sum_sq_slots += start**2
                count_slots += 1
                if day >= 5: weekend_days += 1
                
                end = min(start + mask_len, num_slots)
                real_len = end - start
                if real_len > 0:
                    current_coverage[day, start:end] += real_mask[:real_len]
            
            if cons_types[i, day] == CONS_TYPE_SOFT and start >= 0:
                target_slot = cons_vals[i, day]
                if start != target_slot:
                    soft_pref_penalty += abs(start - target_slot)

        tgt = target_days_arr[i]
        if days_worked_count != tgt:
            contract_penalty += (abs(days_worked_count - tgt) * PENALTY_PER_DAY_MISMATCH)

        if count_slots > 1:
            mean = sum_slots / count_slots
            var = (sum_sq_slots / count_slots) - (mean**2)
            if var > 0: homogeneity_penalty += np.sqrt(var)

        sum_wk_work += weekend_days
        sum_sq_wk_work += weekend_days**2

    if num_emps > 1:
        mean_wk = sum_wk_work / num_emps
        var_wk = (sum_sq_wk_work / num_emps) - (mean_wk**2)
        if var_wk > 0:
            equity_penalty = np.sqrt(var_wk) * W_HOMO * 10.0

    diff = current_coverage - target_demand
    
    under_score = 0.0
    flattened_diff = diff.flatten()
    for val in flattened_diff:
        if val < 0: under_score += abs(val)
    under_score *= W_UNDER
    
    over_score = 0.0
    safe_target = target_demand.flatten()
    for k in range(len(flattened_diff)):
        val = flattened_diff[k]
        if val > 0:
            tgt_val = safe_target[k]
            if tgt_val == 0: tgt_val = 1
            over_score += (val / tgt_val)
    over_score *= (W_OVER * 10.0)

    cost_homo = homogeneity_penalty * W_HOMO
    cost_soft = soft_pref_penalty * W_SOFT
    
    total = under_score + over_score + cost_homo + cost_soft + contract_penalty + equity_penalty
    
    # Ritorna l'array esploso delle loss per i grafici
    return np.array([total, under_score, over_score, cost_homo, cost_soft, contract_penalty, equity_penalty])