Mathesis Wiki

Dies ist eine alte Version des Dokuments!

---

Ich habe mich, um das Thema besser zu verstehen mit den Classic control Acrobot beschäftigt. (https://gym.openai.com/envs/#classic_control)

Hierbei handelt es sich um ein Doppelpendel, welches als Ziel versucht sich über die Linie (höhe 1) zu schwingen.

Die Schwierigkeit liegt bei diesem Environment darin, dass man einen Weg finden muss, dass NN mit den wenigen Erfolgen zu trainieren. (Anfangs kam das Pendel bei 1000 Versuchen ca. 2 mal über die Linie)

Dazu muss man wissen, dass die klassische KI meist nur durch Belohnungen etwas lernt (die sich durch zwischen Etappen erringen lassen).

Zum Vergleich ich habe das Problem mit zwei unterschiedlichen NN getestet (hier Graphen zum Vergleich)

Der Code zum besten Ergebnis ich lass die Trainingsdaten weg, da ich finde dass es am meisten Spaß macht beim lernen zu zugucken.

1. import random
2. import gym
3. import numpy as np
4. from collections import deque
5. from keras.models import Sequential
6. from keras.layers import Dense, Dropout
7. from keras.optimizers import Adam
8. import keras
11. # - - 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 - model = Sequential() - model.add(Dense(16, input_dim=self.state_size, activation='relu', - kernel_regularizer=keras.regularizers.l2(0.00001))) - #model.add(Dropout(0.3)) - #model.add(Dense(24, activation='relu', kernel_regularizer=keras.regularizers.l2(0.00001))) - model.add(Dense(self.action_size, activation='linear')) - model.compile(loss='mse', - optimizer=Adam(lr=self.learning_rate)) - return 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])) - 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 = 7 - - - env = gym.make('Acrobot-v1') - state_size = env.observation_space.shape[0] - 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])#*0.2##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_1000versuche“) - - #