【Pythonコード付】CSVを入力するだけで遺伝的アルゴリズムを実行

import csv
import random
import numpy as np
import pandas as pd

# パラメータ設定
population_size = 50    # 集団のサイズ（初期集団はCSVのデータを使用するので無視）
genome_length = None    # 後でデータの次元から自動設定
crossover_rate = 0.8    # 交叉の確率
mutation_rate = 0.01    # 突然変異の確率
max_generations = 1     # 最大世代数は1
selection_method = "tournament"  # 選択手法: "roulette"、"tournament"、"elite"
crossover_method = "two_point"   # 交叉手法: "one_point"、"two_point"
mutation_method = "gaussian"     # 突然変異手法: "single_bit"、"random_reset"、"gaussian"

# CSVファイルのパスを指定
input_csv = "input_data.csv"   # 入力データCSV（1列目は目的変数（適応度）、2列目以降は説明変数）
output_csv = "next_generation.csv"  # 次世代の出力先

# 選択: ルーレット選択
def roulette_selection(population, fitness_values):
    max_fitness = sum(fitness_values)
    pick = random.uniform(0, max_fitness)
    current = 0
    for ind, fitness_value in zip(population, fitness_values):
        current += fitness_value
        if current > pick:
            return ind

# 選択: トーナメント選択
def tournament_selection(population, fitness_values, k=3):
    selected = random.sample(list(zip(population, fitness_values)), k)
    return max(selected, key=lambda x: x[1])[0]

# 選択: エリート選択
def elite_selection(population, fitness_values, elite_ratio=0.1):
    elite_size = int(len(population) * elite_ratio)
    sorted_population = [x for _, x in sorted(zip(fitness_values, population), reverse=True)]
    return sorted_population[:elite_size]

# 交叉: 1点交叉
def one_point_crossover(parent1, parent2):
    if random.random() < crossover_rate:
        point = random.randint(1, len(parent1) - 1)
        return np.concatenate((parent1[:point], parent2[point:])), np.concatenate((parent2[:point], parent1[point:]))
    return parent1, parent2

# 交叉: 2点交叉
def two_point_crossover(parent1, parent2):
    if random.random() < crossover_rate:
        point1 = random.randint(1, len(parent1) - 2)
        point2 = random.randint(point1 + 1, len(parent1) - 1)
        return (np.concatenate((parent1[:point1], parent2[point1:point2], parent1[point2:])),
                np.concatenate((parent2[:point1], parent1[point1:point2], parent2[point2:])))
    return parent1, parent2

# 突然変異: 単一ビット変異 (バイナリ値)
def single_bit_mutation(individual):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            individual[i] = 1 - individual[i]  # ビットを反転させる
    return individual

# 突然変異: ランダムリセット (連続値用)
def random_reset_mutation(individual, min_max_pairs):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            min_val, max_val = min_max_pairs[i]
            # リセットの際に最小値と最大値の範囲で新しい値を生成
            individual[i] = random.uniform(min_val, max_val)  # int も float も扱える
    return individual

# 突然変異: ガウス突然変異 (連続値)
def gaussian_mutation(individual, mean_std_pairs):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            mean, stddev = mean_std_pairs[i]
            individual[i] += np.random.normal(mean, stddev)
    return individual

# 遺伝的アルゴリズム
def genetic_algorithm(population, fitness_values, min_max_pairs, mean_std_pairs, selection_method, crossover_method, mutation_method):
    next_generation = []

    # エリート選択
    if selection_method == "elite":
        elites = elite_selection(population, fitness_values)
        next_generation.extend(elites)

    # 交叉と突然変異
    while len(next_generation) < len(population):
        # 選択メソッド選択
        if selection_method == "roulette":
            parent1 = roulette_selection(population, fitness_values)
            parent2 = roulette_selection(population, fitness_values)
        elif selection_method == "tournament":
            parent1 = tournament_selection(population, fitness_values)
            parent2 = tournament_selection(population, fitness_values)
        elif selection_method == "elite":
            parent1 = random.choice(elites)
            parent2 = random.choice(elites)

        # 交叉メソッド選択
        if crossover_method == "one_point":
            offspring1, offspring2 = one_point_crossover(parent1, parent2)
        elif crossover_method == "two_point":
            offspring1, offspring2 = two_point_crossover(parent1, parent2)

        # 突然変異メソッド選択
        if mutation_method == "single_bit":
            offspring1 = single_bit_mutation(offspring1)
            offspring2 = single_bit_mutation(offspring2)
        elif mutation_method == "random_reset":
            offspring1 = random_reset_mutation(offspring1, min_max_pairs)
            offspring2 = random_reset_mutation(offspring2, min_max_pairs)
        elif mutation_method == "gaussian":
            offspring1 = gaussian_mutation(offspring1, mean_std_pairs)
            offspring2 = gaussian_mutation(offspring2, mean_std_pairs)
            
        next_generation.append(offspring1)
        next_generation.append(offspring2)

    return next_generation[:len(population)]

# CSVファイルを読み込んでGAを実行
def run_ga_from_csv(input_csv, output_csv, selection_method, crossover_method, mutation_method):
    data = pd.read_csv(input_csv)

    # ヘッダーを取得
    header = data.columns.tolist()

    # 説明変数（2列目以降を初期集団として使用）
    x = data.iloc[:, 1:].values  # 説明変数（1列目は適応度で無視）
    fitness_values = data.iloc[:, 0].values  # 1列目を適応度として使用

    global genome_length
    genome_length = x.shape[1]

    # 各列の最小値と最大値を計算（ランダムリセットに使用）
    min_max_pairs = []
    for col in data.iloc[:, 1:].columns:
        min_val = data[col].min()
        max_val = data[col].max()
        min_max_pairs.append((min_val, max_val))
        
    # 各列の平均と標準偏差を計算（連続値のガウス変異に使用）
    mean_std_pairs = []
    for col in data.iloc[:, 1:].columns:
        mean = data[col].mean()
        stddev = data[col].std()     
        mean_std_pairs.append((mean, stddev))
        

    # 遺伝的アルゴリズムを実行
    next_population = genetic_algorithm(x, fitness_values, min_max_pairs, mean_std_pairs, selection_method, crossover_method, mutation_method)

    # 結果をCSVに書き出す（目的変数列は空白）
    result = pd.DataFrame(next_population)
    result.insert(0, '', '')  # 1列目を空白にする
    result.to_csv(output_csv, index=False, header=header)  # input_data.csvのヘッダーを使用

if __name__ == "__main__":
    run_ga_from_csv(input_csv, output_csv, selection_method=selection_method, crossover_method=crossover_method, mutation_method=mutation_method)

# パラメータ設定
population_size = 50    # 集団のサイズ（初期集団はCSVのデータを使用するので無視）
genome_length = None    # 後でデータの次元から自動設定
crossover_rate = 0.8    # 交叉の確率
mutation_rate = 0.01    # 突然変異の確率
max_generations = 1     # 最大世代数は1
selection_method = "tournament"  # 選択手法: "roulette"、"tournament"、"elite"
crossover_method = "two_point"   # 交叉手法: "one_point"、"two_point"
mutation_method = "single_bit"     # 突然変異手法: "single_bit"、"random_reset"、"gaussian"

# CSVファイルのパスを指定
input_csv = "input_data.csv"   # 入力データCSV（1列目は目的変数（適応度）、2列目以降は説明変数）
output_csv = "next_generation.csv"  # 次世代の出力先

# 選択: ルーレット選択
def roulette_selection(population, fitness_values):
    max_fitness = sum(fitness_values)
    pick = random.uniform(0, max_fitness)
    current = 0
    for ind, fitness_value in zip(population, fitness_values):
        current += fitness_value
        if current > pick:
            return ind

# 選択: トーナメント選択
def tournament_selection(population, fitness_values, k=3):
    selected = random.sample(list(zip(population, fitness_values)), k)
    return max(selected, key=lambda x: x[1])[0]

# 選択: エリート選択
def elite_selection(population, fitness_values, elite_ratio=0.05):
    elite_size = int(len(population) * elite_ratio)
    sorted_population = [x for _, x in sorted(zip(fitness_values, population), reverse=True)]
    return sorted_population[:elite_size]

# 交叉: 1点交叉
def one_point_crossover(parent1, parent2):
    if random.random() < crossover_rate:
        point = random.randint(1, len(parent1) - 1)
        return np.concatenate((parent1[:point], parent2[point:])), np.concatenate((parent2[:point], parent1[point:]))
    return parent1, parent2

# 交叉: 2点交叉
def two_point_crossover(parent1, parent2):
    if random.random() < crossover_rate:
        point1 = random.randint(1, len(parent1) - 2)
        point2 = random.randint(point1 + 1, len(parent1) - 1)
        return (np.concatenate((parent1[:point1], parent2[point1:point2], parent1[point2:])),
                np.concatenate((parent2[:point1], parent1[point1:point2], parent2[point2:])))
    return parent1, parent2

# 突然変異: 単一ビット変異 (バイナリ値)
def single_bit_mutation(individual):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            individual[i] = 1 - individual[i]  # ビットを反転させる
    return individual

# 突然変異: ランダムリセット (連続値用)
def random_reset_mutation(individual, min_max_pairs):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            min_val, max_val = min_max_pairs[i]
            # リセットの際に最小値と最大値の範囲で新しい値を生成
            individual[i] = random.uniform(min_val, max_val)  # int も float も扱える
    return individual

# 突然変異: ガウス突然変異 (連続値)
def gaussian_mutation(individual, mean_std_pairs):
    for i in range(len(individual)):
        if random.random() < mutation_rate:
            mean, stddev = mean_std_pairs[i]
            individual[i] += np.random.normal(mean, stddev)
    return individual

# 遺伝的アルゴリズム
def genetic_algorithm(population, fitness_values, min_max_pairs, mean_std_pairs, selection_method, crossover_method, mutation_method):
    next_generation = []

    # エリート選択
    if selection_method == "elite":
        elites = elite_selection(population, fitness_values)
        next_generation.extend(elites)

    # 交叉と突然変異
    while len(next_generation) < len(population):
        # 選択メソッド選択
        if selection_method == "roulette":
            parent1 = roulette_selection(population, fitness_values)
            parent2 = roulette_selection(population, fitness_values)
        elif selection_method == "tournament":
            parent1 = tournament_selection(population, fitness_values)
            parent2 = tournament_selection(population, fitness_values)
        elif selection_method == "elite":
            parent1 = random.choice(elites)
            parent2 = random.choice(elites)

        # 交叉メソッド選択
        if crossover_method == "one_point":
            offspring1, offspring2 = one_point_crossover(parent1, parent2)
        elif crossover_method == "two_point":
            offspring1, offspring2 = two_point_crossover(parent1, parent2)

        # 突然変異メソッド選択
        if mutation_method == "single_bit":
            offspring1 = single_bit_mutation(offspring1)
            offspring2 = single_bit_mutation(offspring2)
        elif mutation_method == "random_reset":
            offspring1 = random_reset_mutation(offspring1, min_max_pairs)
            offspring2 = random_reset_mutation(offspring2, min_max_pairs)
        elif mutation_method == "gaussian":
            offspring1 = gaussian_mutation(offspring1, mean_std_pairs)
            offspring2 = gaussian_mutation(offspring2, mean_std_pairs)
            
        next_generation.append(offspring1)
        next_generation.append(offspring2)

    return next_generation[:len(population)]

# CSVファイルを読み込んでGAを実行
def run_ga_from_csv(input_csv, output_csv, selection_method, crossover_method, mutation_method):
    data = pd.read_csv(input_csv)

    # ヘッダーを取得
    header = data.columns.tolist()

    # 説明変数（2列目以降を初期集団として使用）
    x = data.iloc[:, 1:].values  # 説明変数（1列目は適応度で無視）
    fitness_values = data.iloc[:, 0].values  # 1列目を適応度として使用

    global genome_length
    genome_length = x.shape[1]

    # 各列の最小値と最大値を計算（ランダムリセットに使用）
    min_max_pairs = []
    for col in data.iloc[:, 1:].columns:
        min_val = data[col].min()
        max_val = data[col].max()
        min_max_pairs.append((min_val, max_val))
        
    # 各列の平均と標準偏差を計算（連続値のガウス変異に使用）
    mean_std_pairs = []
    for col in data.iloc[:, 1:].columns:
        mean = data[col].mean()
        stddev = data[col].std()     
        mean_std_pairs.append((mean, stddev))
        

    # 遺伝的アルゴリズムを実行
    next_population = genetic_algorithm(x, fitness_values, min_max_pairs, mean_std_pairs, selection_method, crossover_method, mutation_method)

    # 結果をCSVに書き出す（目的変数列は空白）
    result = pd.DataFrame(next_population)
    result.insert(0, '', '')  # 1列目を空白にする
    result.to_csv(output_csv, index=False, header=header)  # input_data.csvのヘッダーを使用