|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
import gym import random import numpy as np import matplotlib.pyplot as plt from collections import deque from keras.models import Sequential from keras.optimizers import Adam from keras.layers import Dense, Flatten from keras.layers.convolutional import Conv2D from keras import backend as K env = gym.make('BreakoutDeterministic-v4') observation = env.reset() for i in range(3): # The ball is released after 2 frames if i > 1: print(observation.shape) plt.imshow(observation) plt.show() # Get the next observation observation, _, _, _ = env.step(1) (210, 160, 3) |
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
def preprocess_frame(frame): frame = frame[30:200, 10:150] # grayscale frame and downsize by factor 2 frame = frame[::2, ::2, 0] # set background to 0 frame[frame == 144] = 0 frame[frame == 109] = 0 # set ball and paddles to 1 frame[frame != 0] = 1 return frame obs_preprocessed = preprocess_frame(observation) plt.imshow(obs_preprocessed, cmap='gray') plt.show() (85, 70) |
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 |
class DQLAgent: def __init__(self, cols, rows, n_actions, batch_size=32): self.state_size = (cols, rows, 4) self.n_actions = n_actions self.epsilon = 1. self.epsilon_start, self.epsilon_end = 1.0, 0.1 self.exploration_steps = 1000000. self.epsilon_decay_step = (self.epsilon_start - self.epsilon_end) / self.exploration_steps self.batch_size = batch_size self.discount_factor = 0.99 self.memory = deque(maxlen=400000) self.model = self.build_model() self.target_model = self.build_model() self.optimizer = self.optimizer() self.avg_q_max, self.avg_loss = 0, 0 def optimizer(self): a = K.placeholder(shape=(None,), dtype='int32') y = K.placeholder(shape=(None,), dtype='float32') py_x = self.model.output a_one_hot = K.one_hot(a, self.n_actions) q_value = K.sum(py_x * a_one_hot, axis=1) error = K.abs(y - q_value) quadratic_part = K.clip(error, 0.0, 1.0) linear_part = error - quadratic_part loss = K.mean(0.5 * K.square(quadratic_part) + linear_part) opt = Adam(lr=0.00025, epsilon=0.01) updates = opt.get_updates(self.model.trainable_weights, [], loss) train = K.function([self.model.input, a, y], [loss], updates=updates) return train def build_model(self): model = Sequential() model.add(Conv2D(32, (8, 8), strides=(4, 4), activation='relu', input_shape=self.state_size)) model.add(Conv2D(64, (4, 4), strides=(2, 2), activation='relu')) model.add(Conv2D(64, (3, 3), strides=(1, 1), activation='relu')) model.add(Flatten()) model.add(Dense(512, activation='relu')) model.add(Dense(self.n_actions)) model.summary() return model def update_model(self): self.target_model.set_weights(self.model.get_weights()) def action(self, history): history = np.float32(history / 255.0) if np.random.rand() <= self.epsilon: return random.randrange(self.n_actions) else: q_value = self.model.predict(history) return np.argmax(q_value[0]) def replay(self, history, action, reward, next_history, dead): self.memory.append((history, action, reward, next_history, dead)) def train(self): if len(self.memory) < self.batch_size: return if self.epsilon > self.epsilon_end: self.epsilon -= self.epsilon_decay_step mini_batch = random.sample(self.memory, self.batch_size) history = np.zeros((self.batch_size, self.state_size[0], self.state_size[1], self.state_size[2])) next_history = np.zeros((self.batch_size, self.state_size[0], self.state_size[1], self.state_size[2])) target = np.zeros((self.batch_size,)) action, reward, dead = [], [], [] for i in range(self.batch_size): history[i] = np.float32(mini_batch[i][0] / 255.) next_history[i] = np.float32(mini_batch[i][3] / 255.) action.append(mini_batch[i][1]) reward.append(mini_batch[i][2]) dead.append(mini_batch[i][4]) target_value = self.target_model.predict(next_history) for i in range(self.batch_size): if dead[i]: target[i] = reward[i] else: target[i] = reward[i] + self.discount_factor * \ np.amax(target_value[i]) loss = self.optimizer([history, action, target]) self.avg_loss += loss[0] env = gym.make('BreakoutDeterministic-v4') # General settings n_warmup_steps = 50000 update_model_rate = 10000 cols, rows = 85, 70 n_states = 4 # Hyperparameters batch_size = 32 # Initialization agent = DQLAgent(cols, rows, n_actions=3) scores, episodes = [], [] n_steps = 0 _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d_52 (Conv2D) (None, 20, 16, 32) 8224 _________________________________________________________________ conv2d_53 (Conv2D) (None, 9, 7, 64) 32832 _________________________________________________________________ conv2d_54 (Conv2D) (None, 7, 5, 64) 36928 _________________________________________________________________ flatten_18 (Flatten) (None, 2240) 0 _________________________________________________________________ dense_34 (Dense) (None, 512) 1147392 _________________________________________________________________ dense_35 (Dense) (None, 3) 1539 ================================================================= Total params: 1,226,915 Trainable params: 1,226,915 Non-trainable params: 0 _________________________________________________________________ _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d_55 (Conv2D) (None, 20, 16, 32) 8224 _________________________________________________________________ conv2d_56 (Conv2D) (None, 9, 7, 64) 32832 _________________________________________________________________ conv2d_57 (Conv2D) (None, 7, 5, 64) 36928 _________________________________________________________________ flatten_19 (Flatten) (None, 2240) 0 _________________________________________________________________ dense_36 (Dense) (None, 512) 1147392 _________________________________________________________________ dense_37 (Dense) (None, 3) 1539 ================================================================= Total params: 1,226,915 Trainable params: 1,226,915 Non-trainable params: 0 _________________________________________________________________ while True: done = False dead = False step, score, start_life = 0, 0, 5 observation = env.reset() state = preprocess_frame(observation, cols, rows) history = np.stack((state, state, state, state), axis=2) history = np.reshape([history], (1, cols, rows, n_states)) while not done: # env.render() n_steps += 1 step += 1 # Get action action = agent.action(history) observation, reward, done, info = env.step(action+1) # Extract next state state_next = preprocess_frame(observation, cols, rows) state_next = np.reshape([state_next], (1, cols, rows, 1)) history_next = np.append(state_next, history[:, :, :, :3], axis=3) agent.avg_q_max += np.amax(agent.model.predict(history)[0]) reward = np.clip(reward, -1., 1.) agent.replay(history, action, reward, history_next, dead) agent.train() if n_steps % update_model_rate == 0: agent.update_model() score += reward if dead: dead = False else: history = history_next if done: print('episode {:2d}; score: {:2.0f}; q {:2f}; loss {:2f}; steps {}' .format(n_steps, score, agent.avg_q_max / float(step), agent.avg_loss / float(step), step)) agent.avg_q_max, agent.avg_loss = 0, 0 # Save weights of model if n_steps % 1000 == 0: agent.model.save_weights("breakout_dql.h5") episode 38184; score: 1; q 0.103984; loss 0.003715; steps 184 episode 38344; score: 1; q 0.104243; loss 0.002910; steps 160 ... episode 106467; score: 0; q 0.374253; loss 0.003024; steps 134 env = gym.make('BreakoutDeterministic-v4') agent = DQLAgent(cols, rows, n_action=3) for i in range(5): observation = env.reset() state = pre_processing(observation, cols, rows) history = np.stack((state, state, state, state), axis=2) history = np.reshape([history], (1, cols, rows, n_states)) while not done: env.render() action = agent.get_action(history) observe, reward, done, info = env.step(action+1) |
