Unterschiede

Hier werden die Unterschiede zwischen zwei Versionen gezeigt.

--- ws1819:pacman_code [2019/03/20 18:56]
rhotert
+++ ws1819:pacman_code [2019/03/31 15:30] (aktuell)
rhotert
@@ Zeile 1: / Zeile 1: @@
 **Pacman Code**
+ZIP des Code:  {{:ws1819:pacman_uni.rar|}}
 Wenn ihr das Programm selbst ausprobieren möchtet braucht ihr leider alle Pakete von [[Requirements]], selbst damit ist ein Funktionieren alles andere als Garantiert. Wenn ein Windows update kommt geht wahrscheinlich nichts mehr.
@@ Zeile 16: / Zeile 18: @@
 input_channels = 1
 conv_n_maps = [32, 64, 64]
+conv_kernel_sizes = [(8,8), (4,4), (3,3)]
+conv_strides = [4, 2, 1]
+conv_paddings = ["SAME"] * 3
+#conv_activation = [tf.nn.relu] * 3
+n_hidden_in = 64 * 11 * 10  # conv3 has 64 maps of 11x10 each
+n_hidden = 512
+#hidden_activation = tf.nn.relu
+#n_outputs = env.action_space.n  # 9 discrete actions are available
+#initializer = tf.variance_scaling_initializer()
+class DQNAgent:
+    def __init__(self, state_size, action_size):
+        self.state_size = state_size
+        self.action_size = action_size
+        self.memory = deque(maxlen=2000)
+        self.gamma = 1.0   # discount rate
+        self.epsilon = 1.0  # exploration rate
+        self.epsilon_min = 0.01
+        self.epsilon_decay = 0.999
+        self.learning_rate = 0.001
+        self.model = self._build_model()
+    def _build_model(self):
+        # Einfaches NN
+        vision_model = Sequential()
+        vision_model.add(Conv2D(32, (5, 5) ,activation=None,
+              padding='valid', input_shape=state_size)) ## Achtung hier muss die richtige Dimension rein
+        vision_model.add(keras.layers.advanced_activations.LeakyReLU(alpha=0.05)) #1
+        vision_model.add(MaxPooling2D((2, 2))) #2
+        vision_model.add(Conv2D(64, (3, 3), activation=None, padding='valid')) #3
+        vision_model.add(keras.layers.advanced_activations.LeakyReLU(alpha=0.05)) #4
+        vision_model.add(MaxPooling2D((2, 2))) #5
+        vision_model.add(Flatten()) #6
+    #vision_model.add(keras.layers.core.Dropout(dropout, noise_shape=None, seed=None)) #7
+        vision_model.add(Dense(20,activation=None))#kernel_regularizer=keras.regularizers.l1(reg))) #8
+        vision_model.add(keras.layers.advanced_activations.LeakyReLU(alpha=0.05)) #9
+        vision_model.add(Dense(self.action_size,activation='softmax', name='main_output')) #10
+        vision_model.compile(loss='mse',
+                      optimizer=Adam(lr=self.learning_rate))
+        return vision_model
+    def remember(self, state, action, reward, next_state, done, total, importance):
+        # merkt sich alle bisher durchlaufenen Zustände
+        self.memory.append([state, action, reward, next_state, done,total,importance])
+    def act(self, state):
+        # epsilon-greedy: off-policy oder policy
+        if np.random.rand() <= self.epsilon:
+            return random.randrange(self.action_size)
+        act_values = self.model.predict(state)
+        return np.argmax(act_values[0])  # returns action
+    def replay(self, batch_size):
+        # baut den Vektor der Q-Werte aus
+        # als reward zum Zeitpunkt t + gamma*max(moegliche rewards zum Zeitpunkt t+1)
+        probabilities = np.array([m[-1] for m in self.memory])
+        probabilities = 1./np.sum(probabilities) * probabilities
+        #print( probabilities.shape)
+        minibatch = [self.memory[i] for i in np.random.choice(range(len(self.memory)),size=batch_size, p=probabilities)]
+        states, targets_f = [], []
+        for state, action, reward, next_state, done,total,importance in minibatch:
+            target = reward
+            if not done:
+                target = (reward + self.gamma *
+                          np.amax(self.model.predict(next_state)[0]))
+            #print("Reward: ", reward)
+            target_f = self.model.predict(state)
+            target_f[0][action] = target
+            # Filtering out states and targets for training
+            states.append(state[0])
+            targets_f.append(target_f[0])
+        history = self.model.fit(np.array(states), np.array(targets_f), epochs=1, verbose=0)
+        # Keeping track of loss
+        loss = history.history['loss'][0]
+        if self.epsilon > self.epsilon_min:
+            self.epsilon *= self.epsilon_decay
+        return loss
+    def load(self, name):
+        self.model.load_weights(name)
+    def save(self, name):
+        self.model.save_weights(name)
+EPISODES = 22
+env = gym.make('MsPacman-v0')
+state_size = env.observation_space.shape
+action_size = env.action_space.n
+agent = DQNAgent(state_size, action_size)
+done = False
+batch_size = 32
+for e in range(EPISODES):
+    state = env.reset()
+    state = np.reshape(state, (1,)+ state_size)
+    cum_reward = 0
+    for time in range(500):
+        env.render()
+        action = agent.act(state)
+        next_state, reward, done, _ = env.step(action)
+        #additional_reward = -(state[0,0] + state[0,0]*state[0,2]-state[0,1]*state[0,3])##faktore aus probieren
+        reward = reward #+ additional_reward if not done else 10 #
+        cum_reward += reward
+        next_state = np.reshape(next_state, (1,)+ state_size)
+        agent.remember(state, action, reward, next_state, done,reward,1)
+        state = next_state
+        if done:
+            print("episode: {}/{}, score: {}, e: {:.2}"
+                  .format(e, EPISODES, time, agent.epsilon))
+            break
+        if len(agent.memory) > batch_size:
+            loss = agent.replay(batch_size)
+            # Logging training loss and actual reward every 10 timesteps
+            if time % 10 == 0:
+                print("episode: {}/{}, time: {}, cumulative reward: {:.4f}, loss: {:.4f}".format(e, EPISODES, time, cum_reward, loss))
+    for i in range(time):
+        pos = -i-1
+        agent.memory[-i-2][-2] += reward
+        for j in range(-time,pos):
+            new_total =  agent.memory[j][-2] + agent.memory[pos][2]
+            mem = agent.memory[j]
+            agent.memory[j][-1] =new_total
+    for i in range(time):
+        pos = -i-1
+        imp = max(agent.memory[pos][-2]-agent.model.predict(agent.memory[pos][0])[0,agent.memory[pos][1]],0)
+        mem = agent.memory[pos]
+        agent.memory[pos][-1] = imp
+    agent.save("qlearning_Acrobot_3versuche")
+import gym
+env = gym.make('MsPacman-v0')
+state_size = env.observation_space.shape
+action_size = env.action_space.n
+agent = DQNAgent(state_size, action_size)
+done = False
+batch_size = 32
+zähler=0
+#agent.load("qlearning_Acrobot_3versuche")
+import time  as ti
+for e in range(100):
+    state = env.reset()
+    #state[0] = state[0] + np.random.randn()*0.1
+    #state[1] = state[1] + np.random.randn()*0.1
+    #state[2] = state[2] + np.random.randn()*0.1
+    #state[3] = state[3] + np.random.randn()*0.1
+    #env.env.state = state
+    state = np.reshape(state, [1, state_size])
+    for time in range(2000):
+        env.render()
+        agent.epsilon = 0
+        action = agent.act(state)
+        next_state, reward, done, _ = env.step(action)
+        next_state = np.reshape(next_state, [1, state_size])
+        state = next_state
+        if done:
+            zähler+=1
+            print (zähler,   "Duration: ", time)
+            break
+    else:
+        print ("Volle Zeit")

Mathesis Wiki

Benutzer-Werkzeuge

Webseiten-Werkzeuge

Unterschiede

Seiten-Werkzeuge