TensorFlow Reinforcement Learningアプリケーションの概要

import math import tensorflow as tf from .utils import base_name #      MLP #    class Layer(object): # input_sizes -   ,     MLP  # output_size -   # scope - ,   TensorFlow     #    ( https://www.tensorflow.org/versions/master/how_tos/variable_scope/index.html) def __init__(self, input_sizes, output_size, scope): """Cretes a neural network layer.""" if type(input_sizes) != list: input_sizes = [input_sizes] self.input_sizes = input_sizes self.output_size = output_size self.scope = scope or "Layer" #    with tf.variable_scope(self.scope): #   self.Ws = [] for input_idx, input_size in enumerate(input_sizes): #   W_name = "W_%d" % (input_idx,) #    -   W_initializer = tf.random_uniform_initializer( -1.0 / math.sqrt(input_size), 1.0 / math.sqrt(input_size)) #   -   input_size x output_size W_var = tf.get_variable(W_name, (input_size, output_size), initializer=W_initializer) self.Ws.append(W_var) #      #         self.b = tf.get_variable("b", (output_size,), initializer=tf.constant_initializer(0)) #    # xs -    #     def __call__(self, xs): if type(xs) != list: xs = [xs] assert len(xs) == len(self.Ws), \ "Expected %d input vectors, got %d" % (len(self.Ws), len(xs)) with tf.variable_scope(self.scope): #    #     -  #     -   #  -    +    return sum([tf.matmul(x, W) for x, W in zip(xs, self.Ws)]) + self.b #     #         def variables(self): return [self.b] + self.Ws def copy(self, scope=None): scope = scope or self.scope + "_copy" with tf.variable_scope(scope) as sc: for v in self.variables(): tf.get_variable(base_name(v), v.get_shape(), initializer=lambda x,dtype=tf.float32: v.initialized_value()) sc.reuse_variables() return Layer(self.input_sizes, self.output_size, scope=sc) #   class MLP(object): # input_sizes -    ,   , #       , #         , #      ,      # hiddens -    ,   #  2    100  #    - 4 ,  ,    #   ,      # nonlinearities -     ,     <a href="https://ru.wikipedia.org/wiki/%D0%98%D1%81%D0%BA%D1%83%D1%81%D1%81%D1%82%D0%B2%D0%B5%D0%BD%D0%BD%D1%8B%D0%B9_%D0%BD%D0%B5%D0%B9%D1%80%D0%BE%D0%BD"> </a> # scope - ,   TensorFlow     #    # given_layers -      def __init__(self, input_sizes, hiddens, nonlinearities, scope=None, given_layers=None): self.input_sizes = input_sizes self.hiddens = hiddens self.input_nonlinearity, self.layer_nonlinearities = nonlinearities[0], nonlinearities[1:] self.scope = scope or "MLP" assert len(hiddens) == len(nonlinearities), \ "Number of hiddens must be equal to number of nonlinearities" with tf.variable_scope(self.scope): if given_layers is not None: #    self.input_layer = given_layers[0] self.layers = given_layers[1:] else: #   #    self.input_layer = Layer(input_sizes, hiddens[0], scope="input_layer") self.layers = [] #    for l_idx, (h_from, h_to) in enumerate(zip(hiddens[:-1], hiddens[1:])): self.layers.append(Layer(h_from, h_to, scope="hidden_layer_%d" % (l_idx,))) #   # xs -    #     def __call__(self, xs): if type(xs) != list: xs = [xs] with tf.variable_scope(self.scope): #        hidden = self.input_nonlinearity(self.input_layer(xs)) for layer, nonlinearity in zip(self.layers, self.layer_nonlinearities): #        hidden = nonlinearity(layer(hidden)) return hidden #           def variables(self): res = self.input_layer.variables() for layer in self.layers: res.extend(layer.variables()) return res def copy(self, scope=None): scope = scope or self.scope + "_copy" nonlinearities = [self.input_nonlinearity] + self.layer_nonlinearities given_layers = [self.input_layer.copy()] + [layer.copy() for layer in self.layers] return MLP(self.input_sizes, self.hiddens, nonlinearities, scope=scope, given_layers=given_layers) 

 import numpy as np import random import tensorflow as tf from collections import deque class DiscreteDeepQ(object): #    def __init__(self, observation_size, num_actions, observation_to_actions, optimizer, session, random_action_probability=0.05, exploration_period=1000, store_every_nth=5, train_every_nth=5, minibatch_size=32, discount_rate=0.95, max_experience=30000, target_network_update_rate=0.01, summary_writer=None): #        """Initialized the Deepq object. Based on: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf Parameters ------- observation_size : int length of the vector passed as observation num_actions : int number of actions that the model can execute observation_to_actions: dali model model that implements activate function that can take in observation vector or a batch and returns scores (of unbounded values) for each action for each observation. input shape: [batch_size, observation_size] output shape: [batch_size, num_actions] optimizer: tf.solver.* optimizer for prediction error session: tf.Session session on which to execute the computation random_action_probability: float (0 to 1) exploration_period: int probability of choosing a random action (epsilon form paper) annealed linearly from 1 to random_action_probability over exploration_period store_every_nth: int to further decorrelate samples do not all transitions, but rather every nth transition. For example if store_every_nth is 5, then only 20% of all the transitions is stored. train_every_nth: int normally training_step is invoked every time action is executed. Depending on the setup that might be too often. When this variable is set set to n, then only every n-th time training_step is called will the training procedure actually be executed. minibatch_size: int number of state,action,reward,newstate tuples considered during experience reply dicount_rate: float (0 to 1) how much we care about future rewards. max_experience: int maximum size of the reply buffer target_network_update_rate: float how much to update target network after each iteration. Let's call target_network_update_rate alpha, target network T, and network N. Every time N gets updated we execute: T = (1-alpha)*T + alpha*N summary_writer: tf.train.SummaryWriter writer to log metrics """ """ Deepq  : https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf  ------- observation_size : int     (      ) num_actions : int           observation_to_actions: dali model  (   ),                       :  [batch_size, observation_size]  :  [batch_size, num_actions] optimizer: tf.solver.*           RMSProp session: tf.Session  TensorFlow      random_action_probability: float (0 to 1)   ,              ,      exploration_period: int     ,          1  random_action_probability store_every_nth: int              .            train_every_nth: int  training_step ( )    .   ,    .             minibatch_size: int            RMSProp.      ,  ,     dicount_rate: float (0 to 1)  Q-learning          max_experience: int      target_network_update_rate: float    ,   2  T - target_q_network         N - q_network     ,         .  T       N.     N,    : alpha = target_network_update_rate T = (1-alpha)*T + alpha*N summary_writer: tf.train.SummaryWriter   """ # memorize arguments self.observation_size = observation_size self.num_actions = num_actions self.q_network = observation_to_actions self.optimizer = optimizer self.s = session self.random_action_probability = random_action_probability self.exploration_period = exploration_period self.store_every_nth = store_every_nth self.train_every_nth = train_every_nth self.minibatch_size = minibatch_size self.discount_rate = tf.constant(discount_rate) self.max_experience = max_experience self.target_network_update_rate = \ tf.constant(target_network_update_rate) # deepq state self.actions_executed_so_far = 0 self.experience = deque() self.iteration = 0 self.summary_writer = summary_writer self.number_of_times_store_called = 0 self.number_of_times_train_called = 0 self.create_variables() #     #      # ( ) def linear_annealing(self, n, total, p_initial, p_final): """Linear annealing between p_initial and p_final over total steps - computes value at step n""" if n >= total: return p_final else: return p_initial - (n * (p_initial - p_final)) / (total) #   TensorFlow  #     #   Q-learning def create_variables(self): #   T     N self.target_q_network = self.q_network.copy(scope="target_network") #    # FOR REGULAR ACTION SCORE COMPUTATION with tf.name_scope("taking_action"): #     self.observation = tf.placeholder(tf.float32, (None, self.observation_size), name="observation") #       self.action_scores = tf.identity(self.q_network(self.observation), name="action_scores") tf.histogram_summary("action_scores", self.action_scores) #       self.predicted_actions = tf.argmax(self.action_scores, dimension=1, name="predicted_actions") #    with tf.name_scope("estimating_future_rewards"): # FOR PREDICTING TARGET FUTURE REWARDS #   -   self.next_observation = tf.placeholder(tf.float32, (None, self.observation_size), name="next_observation") #   -    self.next_observation_mask = tf.placeholder(tf.float32, (None,), name="next_observation_mask") #   self.next_action_scores = tf.stop_gradient(self.target_q_network(self.next_observation)) tf.histogram_summary("target_action_scores", self.next_action_scores) #   -  self.rewards = tf.placeholder(tf.float32, (None,), name="rewards") #      target_values = tf.reduce_max(self.next_action_scores, reduction_indices=[1,]) * self.next_observation_mask # r + DF * MAX(Q,s)    Q-learning   self.future_rewards = self.rewards + self.discount_rate * target_values #   N with tf.name_scope("q_value_precition"): # FOR PREDICTION ERROR #         self.action_mask = tf.placeholder(tf.float32, (None, self.num_actions), name="action_mask") #       self.masked_action_scores = tf.reduce_sum(self.action_scores * self.action_mask, reduction_indices=[1,]) #      # - (r + DF * MAX(Q,s) — Q[s',a']) temp_diff = self.masked_action_scores - self.future_rewards #     # RMSProp      self.prediction_error = tf.reduce_mean(tf.square(temp_diff)) #  RMSProp,   -   gradients = self.optimizer.compute_gradients(self.prediction_error) for i, (grad, var) in enumerate(gradients): if grad is not None: gradients[i] = (tf.clip_by_norm(grad, 5), var) # Add histograms for gradients. for grad, var in gradients: tf.histogram_summary(var.name, var) if grad: tf.histogram_summary(var.name + '/gradients', grad) #   -    self.train_op = self.optimizer.apply_gradients(gradients) #       T # T = (1-alpha)*T + alpha*N # UPDATE TARGET NETWORK with tf.name_scope("target_network_update"): self.target_network_update = [] for v_source, v_target in zip(self.q_network.variables(), self.target_q_network.variables()): # this is equivalent to target = (1-alpha) * target + alpha * source update_op = v_target.assign_sub(self.target_network_update_rate * (v_target - v_source)) self.target_network_update.append(update_op) self.target_network_update = tf.group(*self.target_network_update) # summaries tf.scalar_summary("prediction_error", self.prediction_error) self.summarize = tf.merge_all_summaries() self.no_op1 = tf.no_op() #  def action(self, observation): """Given observation returns the action that should be chosen using DeepQ learning strategy. Does not backprop.""" assert len(observation.shape) == 1, \ "Action is performed based on single observation." self.actions_executed_so_far += 1 #     exploration_p = self.linear_annealing(self.actions_executed_so_far, self.exploration_period, 1.0, self.random_action_probability) if random.random() < exploration_p: #   return random.randint(0, self.num_actions - 1) else: #    return self.s.run(self.predicted_actions, {self.observation: observation[np.newaxis,:]})[0] #    #       #    def store(self, observation, action, reward, newobservation): """Store experience, where starting with observation and execution action, we arrived at the newobservation and got thetarget_network_update reward reward If newstate is None, the state/action pair is assumed to be terminal """ if self.number_of_times_store_called % self.store_every_nth == 0: self.experience.append((observation, action, reward, newobservation)) if len(self.experience) > self.max_experience: self.experience.popleft() self.number_of_times_store_called += 1 #   def training_step(self): """Pick a self.minibatch_size exeperiences from reply buffer and backpropage the value function. """ if self.number_of_times_train_called % self.train_every_nth == 0: if len(self.experience) < self.minibatch_size: return #       #   minibatch_size   # sample experience. samples = random.sample(range(len(self.experience)), self.minibatch_size) samples = [self.experience[i] for i in samples] #    #    # bach states states = np.empty((len(samples), self.observation_size)) newstates = np.empty((len(samples), self.observation_size)) action_mask = np.zeros((len(samples), self.num_actions)) newstates_mask = np.empty((len(samples),)) rewards = np.empty((len(samples),)) for i, (state, action, reward, newstate) in enumerate(samples): states[i] = state action_mask[i] = 0 action_mask[i][action] = 1 rewards[i] = reward if newstate is not None: newstates[i] = newstate newstates_mask[i] = 1 else: newstates[i] = 0 newstates_mask[i] = 0 calculate_summaries = self.iteration % 100 == 0 and \ self.summary_writer is not None #   #     #     #    (  #     ) cost, _, summary_str = self.s.run([ self.prediction_error, self.train_op, self.summarize if calculate_summaries else self.no_op1, ], { self.observation: states, self.next_observation: newstates, self.next_observation_mask: newstates_mask, self.action_mask: action_mask, self.rewards: rewards, }) #    self.s.run(self.target_network_update) if calculate_summaries: self.summary_writer.add_summary(summary_str, self.iteration) self.iteration += 1 self.number_of_times_train_called += 1 

 import math import matplotlib.pyplot as plt import numpy as np import random import time from collections import defaultdict from euclid import Circle, Point2, Vector2, LineSegment2 import tf_rl.utils.svg as svg #   #     #     #     class GameObject(object): def __init__(self, position, speed, obj_type, settings): """Esentially represents circles of different kinds, which have position and speed.""" self.settings = settings self.radius = self.settings["object_radius"] self.obj_type = obj_type self.position = position self.speed = speed self.bounciness = 1.0 def wall_collisions(self): """Update speed upon collision with the wall.""" world_size = self.settings["world_size"] for dim in range(2): if self.position[dim] - self.radius <= 0 and self.speed[dim] < 0: self.speed[dim] = - self.speed[dim] * self.bounciness elif self.position[dim] + self.radius + 1 >= world_size[dim] and self.speed[dim] > 0: self.speed[dim] = - self.speed[dim] * self.bounciness def move(self, dt): """Move as if dt seconds passed""" self.position += dt * self.speed self.position = Point2(*self.position) def step(self, dt): """Move and bounce of walls.""" self.wall_collisions() self.move(dt) def as_circle(self): return Circle(self.position, float(self.radius)) def draw(self): """Return svg object for this item.""" color = self.settings["colors"][self.obj_type] return svg.Circle(self.position + Point2(10, 10), self.radius, color=color) # .     # ,    , #    ,   # .      # ,   ^  ,   , #  ,    # observe,      #      class KarpathyGame(object): def __init__(self, settings): """Initiallize game simulator with settings""" self.settings = settings self.size = self.settings["world_size"] self.walls = [LineSegment2(Point2(0,0), Point2(0,self.size[1])), LineSegment2(Point2(0,self.size[1]), Point2(self.size[0], self.size[1])), LineSegment2(Point2(self.size[0], self.size[1]), Point2(self.size[0], 0)), LineSegment2(Point2(self.size[0], 0), Point2(0,0))] self.hero = GameObject(Point2(*self.settings["hero_initial_position"]), Vector2(*self.settings["hero_initial_speed"]), "hero", self.settings) if not self.settings["hero_bounces_off_walls"]: self.hero.bounciness = 0.0 self.objects = [] for obj_type, number in settings["num_objects"].items(): for _ in range(number): self.spawn_object(obj_type) self.observation_lines = self.generate_observation_lines() self.object_reward = 0 self.collected_rewards = [] #        #        # every observation_line sees one of objects or wall and # two numbers representing speed of the object (if applicable) self.eye_observation_size = len(self.settings["objects"]) + 3 # ,  ,    #   -    # additionally there are two numbers representing agents own speed. self.observation_size = self.eye_observation_size * len(self.observation_lines) + 2 self.last_observation = np.zeros(self.observation_size) self.directions = [Vector2(*d) for d in [[1,0], [0,1], [-1,0],[0,-1]]] self.num_actions = len(self.directions) self.objects_eaten = defaultdict(lambda: 0) def perform_action(self, action_id): """Change speed to one of hero vectors""" assert 0 <= action_id < self.num_actions self.hero.speed *= 0.8 self.hero.speed += self.directions[action_id] * self.settings["delta_v"] def spawn_object(self, obj_type): """Spawn object of a given type and add it to the objects array""" radius = self.settings["object_radius"] position = np.random.uniform([radius, radius], np.array(self.size) - radius) position = Point2(float(position[0]), float(position[1])) max_speed = np.array(self.settings["maximum_speed"]) speed = np.random.uniform(-max_speed, max_speed).astype(float) speed = Vector2(float(speed[0]), float(speed[1])) self.objects.append(GameObject(position, speed, obj_type, self.settings)) def step(self, dt): """Simulate all the objects for a given ammount of time. Also resolve collisions with the hero""" for obj in self.objects + [self.hero] : obj.step(dt) self.resolve_collisions() def squared_distance(self, p1, p2): return (p1[0] - p2[0]) ** 2 + (p1[1] - p2[1]) ** 2 def resolve_collisions(self): """If hero touches, hero eats. Also reward gets updated.""" collision_distance = 2 * self.settings["object_radius"] collision_distance2 = collision_distance ** 2 to_remove = [] for obj in self.objects: if self.squared_distance(self.hero.position, obj.position) < collision_distance2: to_remove.append(obj) for obj in to_remove: self.objects.remove(obj) self.objects_eaten[obj.obj_type] += 1 self.object_reward += self.settings["object_reward"][obj.obj_type] self.spawn_object(obj.obj_type) def inside_walls(self, point): """Check if the point is inside the walls""" EPS = 1e-4 return (EPS <= point[0] < self.size[0] - EPS and EPS <= point[1] < self.size[1] - EPS) #    def observe(self): """Return observation vector. For all the observation directions it returns representation of the closest object to the hero - might be nothing, another object or a wall. Representation of observation for all the directions will be concatenated. """ num_obj_types = len(self.settings["objects"]) + 1 # and wall max_speed_x, max_speed_y = self.settings["maximum_speed"] #   observable_distance = self.settings["observation_line_length"] #       relevant_objects = [obj for obj in self.objects if obj.position.distance(self.hero.position) < observable_distance] #     #   # objects sorted from closest to furthest relevant_objects.sort(key=lambda x: x.position.distance(self.hero.position)) observation = np.zeros(self.observation_size) observation_offset = 0 #     for i, observation_line in enumerate(self.observation_lines): # shift to hero position observation_line = LineSegment2(self.hero.position + Vector2(*observation_line.p1), self.hero.position + Vector2(*observation_line.p2)) observed_object = None #      # if end of observation line is outside of walls, we see the wall. if not self.inside_walls(observation_line.p2): observed_object = "**wall**" #      for obj in relevant_objects: if observation_line.distance(obj.position) < self.settings["object_radius"]: #   observed_object = obj break #    # ,      object_type_id = None speed_x, speed_y = 0, 0 proximity = 0 if observed_object == "**wall**": # wall seen #   object_type_id = num_obj_types - 1 #      #  ,  , #  ,  ,  #    #   #    # a wall has fairly low speed... # speed_x, speed_y = 0, 0 # I think relative speed is better than absolute speed_x, speed_y = tuple (-self.hero.speed) # best candidate is intersection between # observation_line and a wall, that's # closest to the hero best_candidate = None for wall in self.walls: candidate = observation_line.intersect(wall) if candidate is not None: if (best_candidate is None or best_candidate.distance(self.hero.position) > candidate.distance(self.hero.position)): best_candidate = candidate if best_candidate is None: # assume it is due to rounding errors # and wall is barely touching observation line proximity = observable_distance else: proximity = best_candidate.distance(self.hero.position) elif observed_object is not None: # agent seen #   #   object_type_id = self.settings["objects"].index(observed_object.obj_type) #       #   speed_x, speed_y = tuple(observed_object.speed - self.hero.speed) intersection_segment = obj.as_circle().intersect(observation_line) assert intersection_segment is not None #     try: proximity = min(intersection_segment.p1.distance(self.hero.position), intersection_segment.p2.distance(self.hero.position)) except AttributeError: proximity = observable_distance for object_type_idx_loop in range(num_obj_types): #  1.0      #    observation[observation_offset + object_type_idx_loop] = 1.0 if object_type_id is not None: #         #     0.0  1.0 #      observation[observation_offset + object_type_id] = proximity / observable_distance #    observation[observation_offset + num_obj_types] = speed_x / max_speed_x observation[observation_offset + num_obj_types + 1] = speed_y / max_speed_y assert num_obj_types + 2 == self.eye_observation_size observation_offset += self.eye_observation_size #       #     observation[observation_offset] = self.hero.speed[0] / max_speed_x observation[observation_offset + 1] = self.hero.speed[1] / max_speed_y assert observation_offset + 2 == self.observation_size self.last_observation = observation return observation def distance_to_walls(self): """Returns distance of a hero to walls""" res = float('inf') for wall in self.walls: res = min(res, self.hero.position.distance(wall)) return res - self.settings["object_radius"] def collect_reward(self): """Return accumulated object eating score + current distance to walls score""" wall_reward = self.settings["wall_distance_penalty"] * \ np.exp(-self.distance_to_walls() / self.settings["tolerable_distance_to_wall"]) assert wall_reward < 1e-3, "You are rewarding hero for being close to the wall!" total_reward = wall_reward + self.object_reward self.object_reward = 0 self.collected_rewards.append(total_reward) return total_reward def plot_reward(self, smoothing = 30): """Plot evolution of reward over time.""" plottable = self.collected_rewards[:] while len(plottable) > 1000: for i in range(0, len(plottable) - 1, 2): plottable[i//2] = (plottable[i] + plottable[i+1]) / 2 plottable = plottable[:(len(plottable) // 2)] x = [] for i in range(smoothing, len(plottable)): chunk = plottable[i-smoothing:i] x.append(sum(chunk) / len(chunk)) plt.plot(list(range(len(x))), x) def generate_observation_lines(self): """Generate observation segments in settings["num_observation_lines"] directions""" result = [] start = Point2(0.0, 0.0) end = Point2(self.settings["observation_line_length"], self.settings["observation_line_length"]) for angle in np.linspace(0, 2*np.pi, self.settings["num_observation_lines"], endpoint=False): rotation = Point2(math.cos(angle), math.sin(angle)) current_start = Point2(start[0] * rotation[0], start[1] * rotation[1]) current_end = Point2(end[0] * rotation[0], end[1] * rotation[1]) result.append( LineSegment2(current_start, current_end)) return result def _repr_html_(self): return self.to_html() def to_html(self, stats=[]): """Return svg representation of the simulator""" stats = stats[:] recent_reward = self.collected_rewards[-100:] + [0] objects_eaten_str = ', '.join(["%s: %s" % (o,c) for o,c in self.objects_eaten.items()]) stats.extend([ "nearest wall = %.1f" % (self.distance_to_walls(),), "reward = %.1f" % (sum(recent_reward)/len(recent_reward),), "objects eaten => %s" % (objects_eaten_str,), ]) scene = svg.Scene((self.size[0] + 20, self.size[1] + 20 + 20 * len(stats))) scene.add(svg.Rectangle((10, 10), self.size)) num_obj_types = len(self.settings["objects"]) + 1 # and wall observation_offset = 0; for line in self.observation_lines: # getting color of the line linecolor = 'black'; linewidth = '1px'; for object_type_idx_loop in range(num_obj_types): if self.last_observation[observation_offset + object_type_idx_loop] < 1.0: if object_type_idx_loop < num_obj_types - 1: linecolor = self.settings["colors"][self.settings["objects"][object_type_idx_loop]]; linewidth = '3px'; observation_offset += self.eye_observation_size scene.add(svg.Line(line.p1 + self.hero.position + Point2(10,10), line.p2 + self.hero.position + Point2(10,10), color = linecolor, stroke = linecolor, stroke_width = linewidth)) for obj in self.objects + [self.hero] : scene.add(obj.draw()) offset = self.size[1] + 15 for txt in stats: scene.add(svg.Text((10, offset + 20), txt, 15)) offset += 20 return scene