Merge pull request rlcode#68 from jcwleo/master

jcwleo · web-flow · commit 2fe6984da684 · 2017-11-28T17:19:34.000+09:00
Add DQN with PER
diff --git a/2-cartpole/1-dqn/SumTree.py b/2-cartpole/1-dqn/SumTree.py
@@ -0,0 +1,55 @@
+import numpy
+
+
+class SumTree:
+    write = 0
+
+    def __init__(self, capacity):
+        self.capacity = capacity
+        self.tree = numpy.zeros(2 * capacity - 1)
+        self.data = numpy.zeros(capacity, dtype=object)
+
+    def _propagate(self, idx, change):
+        parent = (idx - 1) // 2
+
+        self.tree[parent] += change
+
+        if parent != 0:
+            self._propagate(parent, change)
+
+    def _retrieve(self, idx, s):
+        left = 2 * idx + 1
+        right = left + 1
+
+        if left >= len(self.tree):
+            return idx
+
+        if s <= self.tree[left]:
+            return self._retrieve(left, s)
+        else:
+            return self._retrieve(right, s - self.tree[left])
+
+    def total(self):
+        return self.tree[0]
+
+    def add(self, p, data):
+        idx = self.write + self.capacity - 1
+
+        self.data[self.write] = data
+        self.update(idx, p)
+
+        self.write += 1
+        if self.write >= self.capacity:
+            self.write = 0
+
+    def update(self, idx, p):
+        change = p - self.tree[idx]
+
+        self.tree[idx] = p
+        self._propagate(idx, change)
+
+    def get(self, s):
+        idx = self._retrieve(0, s)
+        dataIdx = idx - self.capacity + 1
+
+        return (idx, self.tree[idx], self.data[dataIdx])
diff --git a/2-cartpole/1-dqn/cartpole_only_per.py b/2-cartpole/1-dqn/cartpole_only_per.py
@@ -0,0 +1,224 @@
+import sys
+import gym
+import pylab
+import random
+import numpy as np
+from SumTree import SumTree
+from collections import deque
+from keras.layers import Dense
+from keras.optimizers import Adam
+from keras.models import Sequential
+
+EPISODES = 300
+
+
+# 카트폴 예제에서의 DQN 에이전트
+class DQNAgent:
+    def __init__(self, state_size, action_size):
+        self.render = False
+        self.load_model = False
+
+        # 상태와 행동의 크기 정의
+        self.state_size = state_size
+        self.action_size = action_size
+
+        # DQN 하이퍼파라미터
+        self.discount_factor = 0.99
+        self.learning_rate = 0.001
+        self.epsilon = 1.0
+        self.epsilon_decay = 0.999
+        self.epsilon_min = 0.01
+        self.batch_size = 64
+        self.train_start = 2000
+        self.memory_size = 2000
+
+        # 리플레이 메모리, 최대 크기 2000
+        self.memory = Memory(self.memory_size)
+
+        # 모델과 타깃 모델 생성
+        self.model = self.build_model()
+        self.target_model = self.build_model()
+
+        # 타깃 모델 초기화
+        self.update_target_model()
+
+        if self.load_model:
+            self.model.load_weights("./save_model/cartpole_dqn_trained.h5")
+
+    # 상태가 입력, 큐함수가 출력인 인공신경망 생성
+    def build_model(self):
+        model = Sequential()
+        model.add(Dense(24, input_dim=self.state_size, activation='relu',
+                        kernel_initializer='he_uniform'))
+        model.add(Dense(24, activation='relu',
+                        kernel_initializer='he_uniform'))
+        model.add(Dense(self.action_size, activation='linear',
+                        kernel_initializer='he_uniform'))
+        model.summary()
+        model.compile(loss='mse', optimizer=Adam(lr=self.learning_rate))
+        return model
+
+    # 타깃 모델을 모델의 가중치로 업데이트
+    def update_target_model(self):
+        self.target_model.set_weights(self.model.get_weights())
+
+    # 입실론 탐욕 정책으로 행동 선택
+    def get_action(self, state):
+        if np.random.rand() <= self.epsilon:
+            return random.randrange(self.action_size)
+        else:
+            q_value = self.model.predict(state)
+            return np.argmax(q_value[0])
+
+    # 샘플 <s, a, r, s'>을 리플레이 메모리에 저장
+    def append_sample(self, state, action, reward, next_state, done):
+        if self.epsilon == 1:
+            done = True
+
+        # TD-error 를 구해서 같이 메모리에 저장
+        target = self.model.predict([state])
+        old_val = target[0][action]
+        target_val = self.target_model.predict([next_state])
+        if done:
+            target[0][action] = reward
+        else:
+            target[0][action] = reward + self.discount_factor * (
+                np.amax(target_val[0]))
+        error = abs(old_val - target[0][action])
+
+        self.memory.add(error, (state, action, reward, next_state, done))
+
+    # 리플레이 메모리에서 무작위로 추출한 배치로 모델 학습
+    def train_model(self):
+        if self.epsilon > self.epsilon_min:
+            self.epsilon *= self.epsilon_decay
+
+        # 메모리에서 배치 크기만큼 무작위로 샘플 추출
+        mini_batch = self.memory.sample(self.batch_size)
+
+        errors = np.zeros(self.batch_size)
+        states = np.zeros((self.batch_size, self.state_size))
+        next_states = np.zeros((self.batch_size, self.state_size))
+        actions, rewards, dones = [], [], []
+
+        for i in range(self.batch_size):
+            states[i] = mini_batch[i][1][0]
+            actions.append(mini_batch[i][1][1])
+            rewards.append(mini_batch[i][1][2])
+            next_states[i] = mini_batch[i][1][3]
+            dones.append(mini_batch[i][1][4])
+
+        # 현재 상태에 대한 모델의 큐함수
+        # 다음 상태에 대한 타깃 모델의 큐함수
+        target = self.model.predict(states)
+        target_val = self.target_model.predict(next_states)
+
+        # 벨만 최적 방정식을 이용한 업데이트 타깃
+        for i in range(self.batch_size):
+            old_val = target[i][actions[i]]
+            if dones[i]:
+                target[i][actions[i]] = rewards[i]
+            else:
+                target[i][actions[i]] = rewards[i] + self.discount_factor * (
+                    np.amax(target_val[i]))
+            # TD-error를 저장
+            errors[i] = abs(old_val - target[i][actions[i]])
+
+        # TD-error로 priority 업데이트
+        for i in range(self.batch_size):
+            idx = mini_batch[i][0]
+            self.memory.update(idx, errors[i])
+
+        self.model.fit(states, target, batch_size=self.batch_size,
+                       epochs=1, verbose=0)
+
+
+class Memory:  # stored as ( s, a, r, s_ ) in SumTree
+    e = 0.01
+    a = 0.6
+
+    def __init__(self, capacity):
+        self.tree = SumTree(capacity)
+
+    def _getPriority(self, error):
+        return (error + self.e) ** self.a
+
+    def add(self, error, sample):
+        p = self._getPriority(error)
+        self.tree.add(p, sample)
+
+    def sample(self, n):
+        batch = []
+        segment = self.tree.total() / n
+
+        for i in range(n):
+            a = segment * i
+            b = segment * (i + 1)
+
+            s = random.uniform(a, b)
+            (idx, p, data) = self.tree.get(s)
+            batch.append((idx, data))
+
+        return batch
+
+    def update(self, idx, error):
+        p = self._getPriority(error)
+        self.tree.update(idx, p)
+
+
+if __name__ == "__main__":
+    # CartPole-v1 환경, 최대 타임스텝 수가 500
+    env = gym.make('CartPole-v1')
+    state_size = env.observation_space.shape[0]
+    action_size = env.action_space.n
+
+    # DQN 에이전트 생성
+    agent = DQNAgent(state_size, action_size)
+
+    scores, episodes = [], []
+
+    step = 0
+    for e in range(EPISODES):
+        done = False
+        score = 0
+        # env 초기화
+        state = env.reset()
+        state = np.reshape(state, [1, state_size])
+
+        while not done:
+            if agent.render:
+                env.render()
+            step += 1
+            # 현재 상태로 행동을 선택
+            action = agent.get_action(state)
+            # 선택한 행동으로 환경에서 한 타임스텝 진행
+            next_state, reward, done, info = env.step(action)
+            next_state = np.reshape(next_state, [1, state_size])
+            # 에피소드가 중간에 끝나면 -100 보상
+            r = reward if not done or score+reward == 500 else -10
+            # 리플레이 메모리에 샘플 <s, a, r, s'> 저장
+            agent.append_sample(state, action, r, next_state, done)
+            # 매 타임스텝마다 학습
+            if step >= agent.train_start:
+                agent.train_model()
+
+            score += reward
+            state = next_state
+
+            if done:
+                # 각 에피소드마다 타깃 모델을 모델의 가중치로 업데이트
+                agent.update_target_model()
+
+#                score = score if score == 500 else score + 100
+                # 에피소드마다 학습 결과 출력
+                scores.append(score)
+                episodes.append(e)
+                pylab.plot(episodes, scores, 'b')
+                pylab.savefig("./save_graph/cartpole_dqn.png")
+                print("episode:", e, "  score:", score, "  memory length:",
+                      step if step <= agent.memory_size else agent.memory_size, "  epsilon:", agent.epsilon)
+
+                # 이전 10개 에피소드의 점수 평균이 490보다 크면 학습 중단
+                if np.mean(scores[-min(10, len(scores)):]) > 490:
+                    agent.model.save_weights("./save_model/cartpole_dqn.h5")
+                    sys.exit()
