def test_nnet_numberline_mdp(n_episodes, exploration_prob=0.9, learning_rate=.0005, target_freeze_period=500): reduce_explore = 0.0001 size = 20. mdp = GridSearchMDP(size) actions = mdp.actions(mdp.startState()) print actions features = T.dvector('features') action = T.lscalar('action') reward = T.dscalar('reward') next_features = T.dvector('next_features') n_vis = 2 # for chain mdp hidden_layer_1 = HiddenLayer(n_vis=n_vis, n_hid=len(actions), layer_name='hidden', activation='tanh') output_layer = OutputLayer(layer_name='out', activation='relu') layers = [hidden_layer_1, output_layer] mlp = QNetwork(layers, discount=mdp.discount(), learning_rate=learning_rate) loss, updates = mlp.get_loss_and_updates(features, action, reward, next_features) train_model = theano.function([ theano.Param(features, default=np.zeros(MAX_FEATURES_TEST)), theano.Param(action, default=0), theano.Param(reward, default=0), theano.Param(next_features, default=np.zeros(MAX_FEATURES_TEST)) ], outputs=loss, updates=updates, mode='FAST_RUN') rewards = [] counter = 0 for episode in xrange(n_episodes): curDiscount = mdp.discount() totalReward = 0 cur_state = mdp.startState() print cur_state while not mdp.isEnd(cur_state): counter += 1 if counter % 1000 == 0: mlp.frozen_layers = copy.deepcopy(mlp.layers) if (counter % 100 == 0): print 'cur_state: {}'.format(cur_state) if random.random() < exploration_prob: action = random.choice(actions) action_index = actions.index(action) else: action_index = T.argmax( mlp.fprop([cur_state[0] / mdp.n, cur_state[1] / mdp.n])).eval() action = actions[action_index] if (counter % 100 == 0): print 'action: {}'.format(action) # realAction = action # if action == 0: realAction = -1 transitions = mdp.succAndProbReward( cur_state, action) # previously realAction) if len(transitions) == 0: break # Choose a random transition i = sample([prob for newState, prob, reward in transitions]) newState, prob, reward = transitions[i] #print 'newState: {}'.format(newState) #print 'reward: {}'.format(reward) #print [(p.eval(), p.name) for p in mlp.get_params()] #print [(p.eval(), p.name) for p in mlp.get_params(freeze=True)] #print '\n' reward *= curDiscount totalReward += reward curDiscount *= mdp.discount() loss = train_model([cur_state[0] / mdp.n, cur_state[1] / mdp.n], action_index, reward, [newState[0] / mdp.n, newState[1] / mdp.n ]) # previously action cur_state = newState exploration_prob -= reduce_explore if (exploration_prob < 0.25): exploration_prob = 0.25 rewards.append(totalReward) print('*' * 30) print('episode: {} ended with score: {}'.format(episode, rewards[-1])) print('avg reward: {}'.format(np.mean(rewards[-25:]))) print('explore: {}'.format(exploration_prob)) print('*' * 30) print('\n') return rewards
def simulate_symbolic_online_RL_algorithm(mdp, num_episodes, max_iterations): real_actions = mdp.actions(None) actions = np.arange(len(real_actions)) # these theano variables are used to define the symbolic input of the network features = T.dvector('features') action = T.lscalar('action') reward = T.dscalar('reward') next_features = T.dvector('next_features') learning_rate_symbol = T.dscalar('learning_rate') h1 = HiddenLayer(n_vis=INPUT_DIM, n_hid=HIDDEN_DIM, layer_name='h1') h2 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h2') h3 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h3') h4 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=OUTPUT_DIM, layer_name='h4') # h5 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h5') # h6 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=OUTPUT_DIM, layer_name='h6') layers = [h1, h2, h3, h4] #, h3, h4, h5, h6] learning_rate = 1e-2 explorationProb = .4 regularization_weight = 1e-5 momentum_rate = 9e-1 qnetwork = QNetwork(layers, discount=mdp.discount, momentum_rate=momentum_rate, regularization_weight=regularization_weight) exploration_reduction = (explorationProb - MIN_EXPLORATION_PROB) / num_episodes learning_rate_reduction = (learning_rate - MIN_LEARNING_RATE) / num_episodes # this call gets the symbolic output of the network along with the parameter updates loss, updates = qnetwork.get_loss_and_updates(features, action, reward, next_features, learning_rate_symbol) print 'Building Training Function...' # this defines the theano symbolic function used to train the network # 1st argument is a list of inputs, here the symbolic variables above # 2nd argument is the symbolic output expected # 3rd argument is the dictionary of parameter updates # 4th argument is the compilation mode train_model = theano.function( [theano.Param(features, default=np.zeros(INPUT_DIM)), theano.Param(action, default=0), theano.Param(reward, default=0), theano.Param(next_features, default=np.zeros(HIDDEN_DIM)), learning_rate_symbol], outputs=loss, updates=updates, mode='FAST_RUN') get_action = theano.function([features], qnetwork.get_action(features)) total_rewards = [] total_losses = [] weight_magnitudes = [] print 'Starting Training...' replay_mem = replay_memory.ReplayMemory() for episode in xrange(num_episodes): state = np.array(mdp.start_state) total_reward = 0 total_loss = 0 for iteration in xrange(max_iterations): if random.random() < explorationProb: action = random.choice(actions) else: action = get_action(state) real_action = real_actions[action] transitions = mdp.succAndProbReward(state, real_action) if len(transitions) == 0: # loss += train_model(state, action, 0, next_features) break # Choose a random transition i = sample([prob for newState, prob, reward in transitions]) newState, prob, reward = transitions[i] newState = np.array(newState) sars_tuple = (state, action, np.clip(reward,-1,1), newState) replay_mem.store(sars_tuple) num_samples = 5 if replay_mem.isFull() else 1 for i in range(0, num_samples): random_train_tuple = replay_mem.sample() sample_state = random_train_tuple[0] sample_action = random_train_tuple[1] sample_reward = random_train_tuple[2] sample_new_state = random_train_tuple[3] total_loss += train_model(sample_state, sample_action, sample_reward, sample_new_state, learning_rate) total_reward += reward state = newState explorationProb -= exploration_reduction learning_rate -= learning_rate_reduction total_rewards.append(total_reward * mdp.discount ** iteration) total_losses.append(total_loss) weight_magnitude = qnetwork.get_weight_magnitude() weight_magnitudes.append(weight_magnitude) print 'episode: {}\t\t loss: {}\t\t reward: {}\t\tweight magnitude: {}'.format(episode, round(total_loss, 2), total_reward, weight_magnitude) # return the list of rewards attained return total_rewards, total_losses
def test_nnet_numberline_mdp(n_episodes, exploration_prob=0.9, learning_rate=.0005, target_freeze_period=500): reduce_explore = 0.0001 size = 20. mdp = GridSearchMDP(size) actions = mdp.actions(mdp.startState()) print actions features = T.dvector('features') action = T.lscalar('action') reward = T.dscalar('reward') next_features = T.dvector('next_features') n_vis = 2 # for chain mdp hidden_layer_1 = HiddenLayer(n_vis=n_vis, n_hid=len(actions), layer_name='hidden', activation='tanh') output_layer = OutputLayer(layer_name='out', activation='relu') layers = [hidden_layer_1, output_layer] mlp = QNetwork(layers, discount=mdp.discount(), learning_rate=learning_rate) loss, updates = mlp.get_loss_and_updates(features, action, reward, next_features) train_model = theano.function( [theano.Param(features, default=np.zeros(MAX_FEATURES_TEST)), theano.Param(action, default=0), theano.Param(reward, default=0), theano.Param(next_features, default=np.zeros(MAX_FEATURES_TEST))], outputs=loss, updates=updates, mode='FAST_RUN') rewards = [] counter = 0 for episode in xrange(n_episodes): curDiscount = mdp.discount() totalReward = 0 cur_state = mdp.startState() print cur_state while not mdp.isEnd(cur_state): counter += 1 if counter % 1000 == 0: mlp.frozen_layers = copy.deepcopy(mlp.layers) if (counter % 100 == 0): print 'cur_state: {}'.format(cur_state) if random.random() < exploration_prob: action = random.choice(actions) action_index = actions.index(action) else: action_index = T.argmax(mlp.fprop([cur_state[0]/mdp.n, cur_state[1]/mdp.n])).eval() action = actions[action_index] if (counter % 100 == 0): print 'action: {}'.format(action) # realAction = action # if action == 0: realAction = -1 transitions = mdp.succAndProbReward(cur_state, action) # previously realAction) if len(transitions) == 0: break # Choose a random transition i = sample([prob for newState, prob, reward in transitions]) newState, prob, reward = transitions[i] #print 'newState: {}'.format(newState) #print 'reward: {}'.format(reward) #print [(p.eval(), p.name) for p in mlp.get_params()] #print [(p.eval(), p.name) for p in mlp.get_params(freeze=True)] #print '\n' reward *= curDiscount totalReward += reward curDiscount *= mdp.discount() loss = train_model([cur_state[0]/mdp.n, cur_state[1]/mdp.n], action_index, reward, [newState[0]/mdp.n, newState[1]/mdp.n]) # previously action cur_state = newState exploration_prob -= reduce_explore if (exploration_prob < 0.25): exploration_prob = 0.25 rewards.append(totalReward) print('*' * 30) print('episode: {} ended with score: {}'.format(episode, rewards[-1])) print('avg reward: {}'.format(np.mean(rewards[-25:]))) print('explore: {}'.format(exploration_prob)) print('*' * 30) print('\n') return rewards
def train(gamepath, n_episodes, display_screen, record_weights, reduce_exploration_prob_amount, n_frames_to_skip, exploration_prob, verbose, discount, learning_rate, load_weights, frozen_target_update_period, use_replay_mem): """ :description: trains an agent to play a game :type gamepath: string :param gamepath: path to the binary of the game to be played :type n_episodes: int :param n_episodes: number of episodes of the game on which to train display_screen : whether or not to display the screen of the game record_weights : whether or not to save the weights of the nextwork reduce_exploration_prob_amount : amount to reduce exploration prob each episode to not reduce exploration_prob set to 0 n_frames_to_skip : how frequently to determine a new action to use exploration_prob : probability of choosing a random action verbose : whether or not to print information about the run periodically discount : discount factor used in learning learning_rate : the scaling factor for the sgd update load_weights : whether or not to load weights for the network (set the files directly below) frozen_target_update_period : the number of episodes between reseting the target of the network """ # load the ale interface to interact with ale = ALEInterface() ale.setInt('random_seed', 42) # display/recording settings, doesn't seem to work currently recordings_dir = './recordings/breakout/' # previously "USE_SDL" if display_screen: if sys.platform == 'darwin': import pygame pygame.init() ale.setBool('sound', False) # Sound doesn't work on OSX #ale.setString("record_screen_dir", recordings_dir); elif sys.platform.startswith('linux'): ale.setBool('sound', True) ale.setBool('display_screen', True) ale.loadROM(gamepath) ale.setInt("frame_skip", n_frames_to_skip) # real actions for breakout are [0,1,3,4] real_actions = ale.getMinimalActionSet() # use a list of actions [0,1,2,3] to index into the array of real actions actions = np.arange(len(real_actions)) # these theano variables are used to define the symbolic input of the network features = T.dvector('features') action = T.lscalar('action') reward = T.dscalar('reward') next_features = T.dvector('next_features') # load weights by file name # currently must be loaded by individual hidden layers if load_weights: hidden_layer_1 = file_utils.load_model('weights/hidden0_replay.pkl') hidden_layer_2 = file_utils.load_model('weights/hidden1_replay.pkl') else: # defining the hidden layer network structure # the n_hid of a prior layer must equal the n_vis of a subsequent layer # for q-learning the output layer must be of len(actions) hidden_layer_1 = HiddenLayer(n_vis=NNET_INPUT_DIMENSION, n_hid=NNET_INPUT_DIMENSION, layer_name='hidden1', activation='relu') hidden_layer_2 = HiddenLayer(n_vis=NNET_INPUT_DIMENSION, n_hid=NNET_INPUT_DIMENSION, layer_name='hidden2', activation='relu') hidden_layer_3 = HiddenLayer(n_vis=NNET_INPUT_DIMENSION, n_hid=len(actions), layer_name='hidden3', activation='relu') # the output layer is currently necessary when using tanh units in the # hidden layer in order to prevent a theano warning # currently the relu unit setting of the hidden and output layers is leaky w/ alpha=0.01 output_layer = OutputLayer(layer_name='output', activation='relu') # pass a list of layers to the constructor of the network (here called "mlp") layers = [hidden_layer_1, hidden_layer_2, hidden_layer_3, output_layer] qnetwork = QNetwork(layers, discount=discount, learning_rate=learning_rate) # this call gets the symbolic output of the network # along with the parameter updates expected loss, updates = qnetwork.get_loss_and_updates(features, action, reward, next_features) # this defines the theano symbolic function used to train the network # 1st argument is a list of inputs, here the symbolic variables above # 2nd argument is the symbolic output expected # 3rd argument is the dictionary of parameter updates # 4th argument is the compilation mode train_model = theano.function( [theano.Param(features, default=np.zeros(NNET_INPUT_DIMENSION)), theano.Param(action, default=0), theano.Param(reward, default=0), theano.Param(next_features, default=np.zeros(NNET_INPUT_DIMENSION))], outputs=loss, updates=updates, mode='FAST_RUN') sym_action = qnetwork.get_action(features) get_action = theano.function([features], sym_action) # some containers for collecting information about the training processes rewards = [] losses = [] best_reward = 4 sequence_examples = [] sampled_examples = [] # the preprocessor and feature extractor to use preprocessor = screen_utils.RGBScreenPreprocessor() feature_extractor = feature_extractors.NNetOpenCVBoundingBoxExtractor(max_features=MAX_FEATURES) if use_replay_mem: replay_mem = ReplayMemory() # main training loop, each episode is a full playthrough of the game for episode in xrange(n_episodes): # this implements the frozen target component of the network # by setting the frozen layers of the network to a copy of the current layers if episode % frozen_target_update_period == 0: qnetwork.frozen_layers = copy.deepcopy(qnetwork.layers) # some variables for collecting information about this particular run of the game total_reward = 0 action = 1 counter = 0 reward = 0 loss = 0 previous_param_0 = None # lives here is used for the reward heuristic of subtracting 1 from the reward # when we lose a life. currently commented out this functionality because # i think it might not be helpful. lives = ale.lives() # the initial state of the screen and state screen = np.zeros((preprocessor.dim, preprocessor.dim, preprocessor.channels)) state = { "screen" : screen, "objects" : None, "prev_objects": None, "features": np.zeros(MAX_FEATURES)} # start the actual play through of the game while not ale.game_over(): counter += 1 # get the current features, which is the representation of the state provided to # the "agent" (here just the network directly) features = state["features"] # epsilon greedy action selection (note that exploration_prob is reduced by # reduce_exploration_prob_amount after every game) if random.random() < exploration_prob: action = random.choice(actions) else: # to choose an action from the network, we fprop # the current state and take the argmax of the output # layer (i.e., the action that corresponds to the # maximum q value) action = get_action(features) # take the action and receive the reward reward += ale.act(real_actions[action]) # this is commented out because i think it might not be helpful if ale.lives() < lives: lives = ale.lives() reward -= 1 # get the next screen, preprocess it, initialize the next state next_screen = ale.getScreenRGB() next_screen = preprocessor.preprocess(next_screen) next_state = {"screen": next_screen, "objects": None, "prev_objects": state["objects"]} # get the features for the next state next_features = feature_extractor(next_state, action=None) if use_replay_mem: sars_tuple = (features, action, reward, next_features) replay_mem.store(sars_tuple) num_samples = 5 if replay_mem.isFull() else 1 for i in range(0, num_samples): random_train_tuple = replay_mem.sample() loss += train_model(*random_train_tuple) # collect for pca sequence_examples.append(list(sars_tuple[0]) + [sars_tuple[1]] \ + [sars_tuple[2]] + sars_tuple[3]) sequence_examples = sequence_examples[-100:] sampled_examples.append(list(random_train_tuple[0]) + [random_train_tuple[1]] \ + [random_train_tuple[2]] + random_train_tuple[3]) sampled_examples = sampled_examples[-100:] else: # call the train model function loss += train_model(features, action, reward, next_features) # prepare for the next loop through the game next_state["features"] = next_features state = next_state # weird counter value to avoid interaction with any other counter # loop that might be added, not necessary right now if verbose and counter % PRINT_TRAINING_INFO_PERIOD == 0: print('*' * 15 + ' training information ' + '*' * 15) print('episode: {}'.format(episode)) print('reward: \t{}'.format(reward)) print('avg reward: \t{}'.format(np.mean(rewards))) print 'avg reward (last 25): \t{}'.format(np.mean(rewards[-NUM_EPISODES_AVERAGE_REWARD_OVER:])) print('action: \t{}'.format(real_actions[action])) print('exploration prob: {}'.format(exploration_prob)) param_info = [(p.eval(), p.name) for p in qnetwork.get_params()] for index, (val, name) in enumerate(param_info): if previous_param_0 is None and index == 0: previous_param_0 = val print('parameter {} value: \n{}'.format(name, val)) if index == 0: diff = val - previous_param_0 print('difference from previous param {}: \n{}'.format(name, diff)) print('features: \t{}'.format(features)) print('next_features: \t{}'.format(next_features)) scaled_sequence = preprocessing.scale(np.array(sequence_examples)) scaled_sampled = preprocessing.scale(np.array(sampled_examples)) pca = PCA() _ = pca.fit_transform(scaled_sequence) print('variance explained by first component for sequence: {}%'.format(pca. \ explained_variance_ratio_[0] * 100)) _ = pca.fit_transform(scaled_sampled) print('variance explained by first component for sampled: {}%'.format(pca. \ explained_variance_ratio_[0] * 100)) print('*' * 52) print('\n') # collect info and total reward and also reset the reward to 0 if we reach this point total_reward += reward reward = 0 # collect stats from this game run losses.append(loss) rewards.append(total_reward) # if we got a best reward, inform the user if total_reward > best_reward: best_reward = total_reward print("best reward!: {}".format(total_reward)) # record the weights if record_weights=True # must record the weights of the indiviual layers # only save hidden layers b/c output layer does not have weights if episode != 0 and episode % RECORD_WEIGHTS_PERIOD == 0 and record_weights: file_utils.save_rewards(rewards) file_utils.save_model(qnetwork.layers[0], 'weights/hidden0_{}.pkl'.format(episode)) file_utils.save_model(qnetwork.layers[1], 'weights/hidden1_{}.pkl'.format(episode)) # reduce exploration policy over time if exploration_prob > MINIMUM_EXPLORATION_EPSILON: exploration_prob -= reduce_exploration_prob_amount # inform user of how the episode went and reset the game print('episode: {} ended with score: {}\tloss: {}'.format(episode, rewards[-1], losses[-1])) ale.reset_game() # return the list of rewards attained return rewards
def simulate_symbolic_online_RL_algorithm(mdp, num_episodes, max_iterations): real_actions = mdp.actions(None) actions = np.arange(len(real_actions)) # these theano variables are used to define the symbolic input of the network features = T.dvector('features') action = T.lscalar('action') reward = T.dscalar('reward') next_features = T.dvector('next_features') learning_rate_symbol = T.dscalar('learning_rate') h1 = HiddenLayer(n_vis=INPUT_DIM, n_hid=HIDDEN_DIM, layer_name='h1') h2 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h2') h3 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h3') h4 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=OUTPUT_DIM, layer_name='h4') # h5 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=HIDDEN_DIM, layer_name='h5') # h6 = HiddenLayer(n_vis=HIDDEN_DIM, n_hid=OUTPUT_DIM, layer_name='h6') layers = [h1, h2, h3, h4] #, h3, h4, h5, h6] learning_rate = 1e-2 explorationProb = .4 regularization_weight = 1e-5 momentum_rate = 9e-1 qnetwork = QNetwork(layers, discount=mdp.discount, momentum_rate=momentum_rate, regularization_weight=regularization_weight) exploration_reduction = (explorationProb - MIN_EXPLORATION_PROB) / num_episodes learning_rate_reduction = (learning_rate - MIN_LEARNING_RATE) / num_episodes # this call gets the symbolic output of the network along with the parameter updates loss, updates = qnetwork.get_loss_and_updates(features, action, reward, next_features, learning_rate_symbol) print 'Building Training Function...' # this defines the theano symbolic function used to train the network # 1st argument is a list of inputs, here the symbolic variables above # 2nd argument is the symbolic output expected # 3rd argument is the dictionary of parameter updates # 4th argument is the compilation mode train_model = theano.function([ theano.Param(features, default=np.zeros(INPUT_DIM)), theano.Param(action, default=0), theano.Param(reward, default=0), theano.Param(next_features, default=np.zeros(HIDDEN_DIM)), learning_rate_symbol ], outputs=loss, updates=updates, mode='FAST_RUN') get_action = theano.function([features], qnetwork.get_action(features)) total_rewards = [] total_losses = [] weight_magnitudes = [] print 'Starting Training...' replay_mem = replay_memory.ReplayMemory() for episode in xrange(num_episodes): state = np.array(mdp.start_state) total_reward = 0 total_loss = 0 for iteration in xrange(max_iterations): if random.random() < explorationProb: action = random.choice(actions) else: action = get_action(state) real_action = real_actions[action] transitions = mdp.succAndProbReward(state, real_action) if len(transitions) == 0: # loss += train_model(state, action, 0, next_features) break # Choose a random transition i = sample([prob for newState, prob, reward in transitions]) newState, prob, reward = transitions[i] newState = np.array(newState) sars_tuple = (state, action, np.clip(reward, -1, 1), newState) replay_mem.store(sars_tuple) num_samples = 5 if replay_mem.isFull() else 1 for i in range(0, num_samples): random_train_tuple = replay_mem.sample() sample_state = random_train_tuple[0] sample_action = random_train_tuple[1] sample_reward = random_train_tuple[2] sample_new_state = random_train_tuple[3] total_loss += train_model(sample_state, sample_action, sample_reward, sample_new_state, learning_rate) total_reward += reward state = newState explorationProb -= exploration_reduction learning_rate -= learning_rate_reduction total_rewards.append(total_reward * mdp.discount**iteration) total_losses.append(total_loss) weight_magnitude = qnetwork.get_weight_magnitude() weight_magnitudes.append(weight_magnitude) print 'episode: {}\t\t loss: {}\t\t reward: {}\t\tweight magnitude: {}'.format( episode, round(total_loss, 2), total_reward, weight_magnitude) # return the list of rewards attained return total_rewards, total_losses