Q学習による2次関数の最小値探索

import numpy as np
import matplotlib.pyplot as plt
import random

# ============================================
# 2次関数の係数定義 / Define coefficients of the quadratic function
# f(x) = ax^2 + bx + c
# ============================================
a = 1.0
b = -6.0
c = 8.0

# 2次関数の定義 / Define the quadratic function
def quadratic_function(x):
    return a * x**2 + b * x + c

# ============================================
# 状態空間の定義 / Define the range and resolution of the state space
# ============================================
x_min = -10             # xの最小値 / minimum x value
x_max = 10              # xの最大値 / maximum x value
x_resolution = 0.1      # xの刻み幅 / resolution of x

# 状態リスト（xを離散化）/ Discretize x into states
states = np.arange(x_min, x_max + x_resolution, x_resolution)
n_states = len(states)  # 状態数 / number of states

# ============================================
# 行動空間の定義 / Define action space
# 左（-1）、そのまま（0）、右（+1）/ Move left, stay, move right
# ============================================
actions = [-1, 0, 1]
n_actions = len(actions)

# ============================================
# Qテーブルの初期化 / Initialize Q-table (state × action)
# ============================================
Q = np.zeros((n_states, n_actions))

# ============================================
# Q学習のハイパーパラメータ設定 / Hyperparameters for Q-learning
# ============================================
alpha = 0.1        # 学習率 / learning rate
gamma = 0.9        # 割引率 / discount factor
epsilon = 0.1      # 探索率 / exploration rate
n_episodes = 3000  # 総エピソード数 / number of episodes

# ============================================
# 学習履歴の保存用リスト / Lists for tracking learning progress
# ============================================
history = []         # エピソードごとのベスト結果 / (episode, best_x, f(best_x), total_reward)
total_rewards = []   # エピソードごとの累積報酬 / total reward per episode

# ============================================
# 学習ループ / Q-learning training loop
# ============================================
for episode in range(n_episodes):
    # ランダムに初期状態を選択 / Randomly choose initial state
    state_idx = random.randint(0, n_states - 1)
    total_reward = 0  # 累積報酬の初期化 / initialize cumulative reward

    for step in range(100):
        # ε-greedy による行動選択 / Choose action (exploration vs exploitation)
        if random.random() < epsilon:
            action_idx = random.randint(0, n_actions - 1)  # 探索 / explore
        else:
            action_idx = np.argmax(Q[state_idx])           # 活用 / exploit

        # 次の状態を計算 / Compute next state index
        next_state_idx = state_idx + actions[action_idx]
        if next_state_idx < 0 or next_state_idx >= n_states:
            next_state_idx = state_idx  # 範囲外ならそのまま / stay in range

        # 現在と次のxの値を取得 / Get x and next_x
        x = states[state_idx]
        next_x = states[next_state_idx]

        # 報酬は -f(x)（関数の最小化を目的とするため）/ Reward is -f(x) to encourage minimization
        reward = -quadratic_function(next_x)
        total_reward += reward  # 累積報酬を加算 / accumulate reward

        # Q値の更新 / Q-value update (Bellman equation)
        Q[state_idx, action_idx] += alpha * (
            reward + gamma * np.max(Q[next_state_idx]) - Q[state_idx, action_idx]
        )

        # 状態の更新 / Move to next state
        state_idx = next_state_idx

    # ベストxとそのf(x)を記録 / Record best x and its function value
    best_idx = np.argmax(np.max(Q, axis=1))
    best_x = states[best_idx]
    best_y = quadratic_function(best_x)
    history.append((episode, best_x, best_y, total_reward))
    total_rewards.append(total_reward)

    # 100エピソードごとに進捗表示 / Show progress every 100 episodes
    if (episode + 1) % 100 == 0 or episode == 0:
        print(f"Episode {episode + 1}/{n_episodes}")
        print(f"  → Best x so far: {best_x:.4f}")
        print(f"  → f(x): {best_y:.4f}")
        print(f"  → Total reward in this episode: {total_reward:.2f}")
        print("-" * 40)

# ============================================
# 最終結果の表示 / Show final result
# ============================================
best_state_idx = np.argmax(np.max(Q, axis=1))
best_x = states[best_state_idx]
best_y = quadratic_function(best_x)

print("\nFinal Result")
print(f"x = {best_x:.4f} gives minimum f(x) = {best_y:.4f}")

# ============================================
# 関数と最小値のプロット / Plot the function and found minimum
# ============================================
x_vals = np.linspace(x_min, x_max, 500)
y_vals = quadratic_function(x_vals)

plt.figure(figsize=(10, 5))
plt.plot(x_vals, y_vals, label="f(x) = ax² + bx + c")
plt.plot(best_x, best_y, 'ro', label="Minimum (via RL)")
plt.xlabel("x")
plt.ylabel("f(x)")
plt.title("Quadratic Function and Minimum Found by Reinforcement Learning")
plt.grid(True)
plt.legend()
plt.show()

# ============================================
# 学習曲線のプロット / Plot learning curve of total reward
# ============================================
plt.figure(figsize=(10, 4))
plt.plot(total_rewards)
plt.xlabel("Episode")
plt.ylabel("Total Reward")
plt.title("Learning Curve: Total Reward per Episode")
plt.grid(True)
plt.show()
結果