Model-based policy networks

Continuing on our reinforcement learning path, we will consider here how to build a “model for our environment” and then use this model to train our policy network, instead of the “actual environment”.

• Simple Reinforcement Learning with Tensorflow: Part 3 - Model-Based RL
• https://github.com/awjuliani/DeepRL-Agents/blob/master/Model-Network.ipynb

• So our reference implementation is here:

  import numpy as np
  import random
  import gym
  
  import tensorflow as tf
  import tensorflow.contrib.slim as slim
  import numpy as np
  
  from nv.core.utils import *
  
  # cf. https://medium.com/@awjuliani/simple-reinforcement-learning-with-tensorflow-part-3-model-based-rl-9a6fe0cce99
  
  def train_model_policy_network():
      logDEBUG("Building environment...")
      env = gym.make('CartPole-v0')
      
      # hyperparameters
      H = 8 # number of hidden layer neurons
      learning_rate = 1e-2
      gamma = 0.99 # discount factor for reward
      decay_rate = 0.99 # decay factor for RMSProp leaky sum of grad^2
      resume = False # resume from previous checkpoint?
  
      model_bs = 3 # Batch size when learning from model
      real_bs = 3 # Batch size when learning from real environment
  
      # model initialization
      D = 4 # input dimensionality
  
      # Policy network:
      tf.reset_default_graph()
      observations = tf.placeholder(tf.float32, [None,4] , name="input_x")
      W1 = tf.get_variable("W1", shape=[4, H],
              initializer=tf.contrib.layers.xavier_initializer())
      layer1 = tf.nn.relu(tf.matmul(observations,W1))
      W2 = tf.get_variable("W2", shape=[H, 1],
              initializer=tf.contrib.layers.xavier_initializer())
      score = tf.matmul(layer1,W2)
      probability = tf.nn.sigmoid(score)
  
      tvars = tf.trainable_variables()
      input_y = tf.placeholder(tf.float32,[None,1], name="input_y")
      advantages = tf.placeholder(tf.float32,name="reward_signal")
      
      adam = tf.train.AdamOptimizer(learning_rate=learning_rate)
      W1Grad = tf.placeholder(tf.float32,name="batch_grad1")
      W2Grad = tf.placeholder(tf.float32,name="batch_grad2")
      batchGrad = [W1Grad,W2Grad]
      loglik = tf.log(input_y*(input_y - probability) + (1 - input_y)*(input_y + probability))
      loss = -tf.reduce_mean(loglik * advantages) 
      newGrads = tf.gradients(loss,tvars)
      updateGrads = adam.apply_gradients(zip(batchGrad,tvars))
  
  
      # Model network:
      mH = 256 # model layer size
  
      input_data = tf.placeholder(tf.float32, [None, 5])
      with tf.variable_scope('rnnlm'):
          softmax_w = tf.get_variable("softmax_w", [mH, 50])
          softmax_b = tf.get_variable("softmax_b", [50])
  
      previous_state = tf.placeholder(tf.float32, [None,5] , name="previous_state")
      W1M = tf.get_variable("W1M", shape=[5, mH],
              initializer=tf.contrib.layers.xavier_initializer())
      B1M = tf.Variable(tf.zeros([mH]),name="B1M")
      layer1M = tf.nn.relu(tf.matmul(previous_state,W1M) + B1M)
      W2M = tf.get_variable("W2M", shape=[mH, mH],
              initializer=tf.contrib.layers.xavier_initializer())
      B2M = tf.Variable(tf.zeros([mH]),name="B2M")
      layer2M = tf.nn.relu(tf.matmul(layer1M,W2M) + B2M)
      wO = tf.get_variable("wO", shape=[mH, 4],
              initializer=tf.contrib.layers.xavier_initializer())
      wR = tf.get_variable("wR", shape=[mH, 1],
              initializer=tf.contrib.layers.xavier_initializer())
      wD = tf.get_variable("wD", shape=[mH, 1],
              initializer=tf.contrib.layers.xavier_initializer())
  
      bO = tf.Variable(tf.zeros([4]),name="bO")
      bR = tf.Variable(tf.zeros([1]),name="bR")
      bD = tf.Variable(tf.ones([1]),name="bD")
  
  
      predicted_observation = tf.matmul(layer2M,wO,name="predicted_observation") + bO
      predicted_reward = tf.matmul(layer2M,wR,name="predicted_reward") + bR
      predicted_done = tf.sigmoid(tf.matmul(layer2M,wD,name="predicted_done") + bD)
  
      true_observation = tf.placeholder(tf.float32,[None,4],name="true_observation")
      true_reward = tf.placeholder(tf.float32,[None,1],name="true_reward")
      true_done = tf.placeholder(tf.float32,[None,1],name="true_done")
  
  
      predicted_state = tf.concat([predicted_observation,predicted_reward,predicted_done],1)
  
      observation_loss = tf.square(true_observation - predicted_observation)
  
      reward_loss = tf.square(true_reward - predicted_reward)
  
      done_loss = tf.multiply(predicted_done, true_done) + tf.multiply(1-predicted_done, 1-true_done)
      done_loss = -tf.log(done_loss)
  
      model_loss = tf.reduce_mean(observation_loss + done_loss + reward_loss)
  
      modelAdam = tf.train.AdamOptimizer(learning_rate=learning_rate)
      updateModel = modelAdam.minimize(model_loss)
  
      # helper functions:
      def resetGradBuffer(gradBuffer):
          for ix,grad in enumerate(gradBuffer):
              gradBuffer[ix] = grad * 0
          return gradBuffer
              
      def discount_rewards(r):
          """ take 1D float array of rewards and compute discounted reward """
          discounted_r = np.zeros_like(r)
          running_add = 0
          for t in reversed(range(0, r.size)):
              running_add = running_add * gamma + r[t]
              discounted_r[t] = running_add
          return discounted_r
  
  
      # This function uses our model to produce a new state when given a previous state and action
      def stepModel(sess, xs, action):
          toFeed = np.reshape(np.hstack([xs[-1][0],np.array(action)]),[1,5])
          myPredict = sess.run([predicted_state],feed_dict={previous_state: toFeed})
          # We should clip the mode reward here:
          reward = myPredict[0][:,4]
          if abs(reward)>1000.0:
              logDEBUG("Clipping model reward value %f" % reward)
              reward = np.clip(reward, -1000.0,1000.0)
  
          observation = myPredict[0][:,0:4]
          observation[:,0] = np.clip(observation[:,0],-2.4,2.4)
          observation[:,2] = np.clip(observation[:,2],-0.4,0.4)
          doneP = np.clip(myPredict[0][:,5],0,1)
          if doneP > 0.1 or len(xs)>= 300:
              done = True
          else:
              done = False
          return observation, reward, done
  
      # Training the model and policy together:
      xs,drs,ys,ds = [],[],[],[]
      running_reward = None
      reward_sum = 0
      episode_number = 1
      real_episodes = 1
      init = tf.global_variables_initializer()
      batch_size = real_bs
  
      drawFromModel = False # When set to True, will use model for observations
      trainTheModel = True # Whether to train the model
      trainThePolicy = False # Whether to train the policy
      switch_point = 1
  
      # Launch the graph
      with tf.Session() as sess:
          rendering = False
          sess.run(init)
          observation = env.reset()
          x = observation
          gradBuffer = sess.run(tvars)
          gradBuffer = resetGradBuffer(gradBuffer)
          
          while episode_number <= 5000:
              # Start displaying environment once performance is acceptably high.
              # No rendering support yet.
              # if (reward_sum/batch_size > 150 and drawFromModel == False) or rendering == True : 
              #     env.render()
              #     rendering = True
                  
              x = np.reshape(observation,[1,4])
  
              tfprob = sess.run(probability,feed_dict={observations: x})
              action = 1 if np.random.uniform() < tfprob else 0
  
              # record various intermediates (needed later for backprop)
              xs.append(x) 
              y = 1 if action == 0 else 0 
              ys.append(y)
              
              # step the  model or real environment and get new measurements
              if drawFromModel == False:
                  observation, reward, done, info = env.step(action)
                  CHECK(reward <= 1000.0, "Invalid reward value from env: %f"%reward)
              else:
                  observation, reward, done = stepModel(sess,xs,action)
                  CHECK(reward <= 1000.0, "Invalid reward value from model: %f"%reward)
                      
              reward_sum += reward
              
              ds.append(done*1)
              drs.append(reward) # record reward (has to be done after we call step() to get reward for previous action)
  
              if done: 
                  
                  if drawFromModel == False: 
                      real_episodes += 1
                  episode_number += 1
  
                  # stack together all inputs, hidden states, action gradients, and rewards for this episode
                  epx = np.vstack(xs)
                  epy = np.vstack(ys)
                  epr = np.vstack(drs)
                  epd = np.vstack(ds)
                  xs,drs,ys,ds = [],[],[],[] # reset array memory
                  
                  if trainTheModel == True:
                      actions = np.array([np.abs(y-1) for y in epy][:-1])
                      state_prevs = epx[:-1,:]
                      state_prevs = np.hstack([state_prevs,actions])
                      state_nexts = epx[1:,:]
                      rewards = np.array(epr[1:,:])
                      dones = np.array(epd[1:,:])
                      state_nextsAll = np.hstack([state_nexts,rewards,dones])
  
                      feed_dict={previous_state: state_prevs, true_observation: state_nexts,true_done:dones,true_reward:rewards}
                      loss,pState,_ = sess.run([model_loss,predicted_state,updateModel],feed_dict)
                  if trainThePolicy == True:
                      discounted_epr = discount_rewards(epr).astype('float32')
                      discounted_epr -= np.mean(discounted_epr)
                      dev = np.std(discounted_epr)
                      if dev > 0.0:
                          discounted_epr /= dev
                      
                      tGrad = sess.run(newGrads,feed_dict={observations: epx, input_y: epy, advantages: discounted_epr})
                      
                      # If gradients becom too large, end training process
                      if np.sum(tGrad[0] == tGrad[0]) == 0:
                          break
                      for ix,grad in enumerate(tGrad):
                          gradBuffer[ix] += grad
                      
                  if switch_point + batch_size == episode_number: 
                      switch_point = episode_number
                      if trainThePolicy == True:
                          sess.run(updateGrads,feed_dict={W1Grad: gradBuffer[0],W2Grad:gradBuffer[1]})
                          gradBuffer = resetGradBuffer(gradBuffer)
  
                      running_reward = reward_sum if running_reward is None else running_reward * 0.99 + reward_sum * 0.01
                      if drawFromModel == False:
                          logDEBUG('World Perf: Episode %d, Reward %f, action: %d, mean reward %f.' % (real_episodes,reward_sum/real_bs,action, running_reward/real_bs))
                          if reward_sum/batch_size >= 200:
                              break
                      reward_sum = 0
  
                      # Once the model has been trained on 100 episodes, we start alternating between training the policy
                      # from the model and training the model from the real environment.
                      if episode_number > 100:
                          drawFromModel = not drawFromModel
                          trainTheModel = not trainTheModel
                          trainThePolicy = not trainThePolicy
                  
                  if drawFromModel == True:
                      observation = np.random.uniform(-0.1,0.1,[4]) # Generate reasonable starting point
                      batch_size = model_bs
                  else:
                      observation = env.reset()
                      batch_size = real_bs
  
      logDEBUG("Done with %d real episodes."% real_episodes)
  

• This implementation works, except that I had to clamp the reward values predicted by the model network to some “reasonable value”:

      def stepModel(sess, xs, action):
          toFeed = np.reshape(np.hstack([xs[-1][0],np.array(action)]),[1,5])
          myPredict = sess.run([predicted_state],feed_dict={previous_state: toFeed})
          # We should clip the mode reward here:
          reward = myPredict[0][:,4]
          if abs(reward)>1000.0:
              logDEBUG("Clipping model reward value %f" % reward)
              reward = np.clip(reward, -1000.0,1000.0)
  

• Because, we currently get very large reward predictions some times, such as:

  2019-03-09T21:46:49.130780 [DEBUG] Clipping model reward value 27099967488.000000
  2019-03-09T21:46:49.132562 [DEBUG] Clipping model reward value 29887655936.000000
  2019-03-09T21:46:49.186213 [DEBUG] World Perf: Episode 187, Reward 34.666667, action: 0, mean reward 1337.982666.
  2019-03-09T21:46:49.268631 [DEBUG] World Perf: Episode 190, Reward 34.333333, action: 1, mean reward 1311.774658.

• And thus, without this clamping, the mean reward value gets completely “out of control”

• Nothing new to say on the policy network: it's exactly the same as the one we used in the post on Full policy Gradient agent
• Let's rather focus on the model network.
• Our input for this network has a dimension D=5
• Actually, the following part of the code is useless (so I will remove it):

      input_data = tf.placeholder(tf.float32, [None, 5])
      with tf.variable_scope('rnnlm'):
          softmax_w = tf.get_variable("softmax_w", [mH, 50])
          softmax_b = tf.get_variable("softmax_b", [50])
  

• For the model, we use 2 fully connected ReLU layers with mH=256 units each
• Then we get 3 different outputs from those layers: the predicted next observation, next reward and next done state.
• Then we compute the loss for each of those outputs, and we sum them before we can do the optimization step.

• ⇒ The clever thing here is really that from a single network, we can consider that we “split the outputs” into conceptually separated components, which could be a very useful trick in general.

• Then for the actual training we train the model and the policy network separately: the model is trained from inputs from the real environement, whereas the policy is trained from inputs from our model network, nothing too fancy here.