View on Run in Google Colab View source on GitHub Download notebook


This example shows how to train a REINFORCE agent on the Cartpole environment using the TF-Agents library, similar to the DQN tutorial.

Cartpole environment

We will walk you through all the components in a Reinforcement Learning (RL) pipeline for training, evaluation and data collection.


If you haven't installed the following dependencies, run:

sudo apt-get install -y xvfb ffmpeg
pip install -q 'gym==0.10.11'
pip install -q 'imageio==2.4.0'
pip install -q PILLOW
pip install -q 'pyglet==1.3.2'
pip install -q pyvirtualdisplay
pip install -q tf-agents

ffmpeg is already the newest version (7:3.4.8-0ubuntu0.2).
xvfb is already the newest version (2:1.19.6-1ubuntu4.4).
The following packages were automatically installed and are no longer required:
  dconf-gsettings-backend dconf-service dkms freeglut3 freeglut3-dev
  glib-networking glib-networking-common glib-networking-services
  gsettings-desktop-schemas libcairo-gobject2 libcolord2 libdconf1
  libegl1-mesa libepoxy0 libglu1-mesa libglu1-mesa-dev libgtk-3-0
  libgtk-3-common libice-dev libjansson4 libjson-glib-1.0-0
  libjson-glib-1.0-common libproxy1v5 librest-0.7-0 libsm-dev
  libsoup-gnome2.4-1 libsoup2.4-1 libxi-dev libxmu-dev libxmu-headers
  libxnvctrl0 libxt-dev linux-gcp-headers-5.0.0-1026
  linux-headers-5.0.0-1026-gcp linux-image-5.0.0-1026-gcp
  linux-modules-5.0.0-1026-gcp pkg-config policykit-1-gnome python3-xkit
  screen-resolution-extra xserver-xorg-core-hwe-18.04
Use 'sudo apt autoremove' to remove them.
0 upgraded, 0 newly installed, 0 to remove and 85 not upgraded.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import base64
import imageio
import IPython
import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import PIL.Image
import pyvirtualdisplay

import tensorflow as tf

from tf_agents.agents.reinforce import reinforce_agent
from tf_agents.drivers import dynamic_step_driver
from tf_agents.environments import suite_gym
from tf_agents.environments import tf_py_environment
from tf_agents.eval import metric_utils
from tf_agents.metrics import tf_metrics
from tf_agents.networks import actor_distribution_network
from tf_agents.replay_buffers import tf_uniform_replay_buffer
from tf_agents.trajectories import trajectory
from tf_agents.utils import common


# Set up a virtual display for rendering OpenAI gym environments.
display = pyvirtualdisplay.Display(visible=0, size=(1400, 900)).start()


env_name = "CartPole-v0" # @param {type:"string"}
num_iterations = 250 # @param {type:"integer"}
collect_episodes_per_iteration = 2 # @param {type:"integer"}
replay_buffer_capacity = 2000 # @param {type:"integer"}

fc_layer_params = (100,)

learning_rate = 1e-3 # @param {type:"number"}
log_interval = 25 # @param {type:"integer"}
num_eval_episodes = 10 # @param {type:"integer"}
eval_interval = 50 # @param {type:"integer"}


Environments in RL represent the task or problem that we are trying to solve. Standard environments can be easily created in TF-Agents using suites. We have different suites for loading environments from sources such as the OpenAI Gym, Atari, DM Control, etc., given a string environment name.

Now let us load the CartPole environment from the OpenAI Gym suite.

env = suite_gym.load(env_name)

We can render this environment to see how it looks. A free-swinging pole is attached to a cart. The goal is to move the cart right or left in order to keep the pole pointing up.



The time_step = environment.step(action) statement takes action in the environment. The TimeStep tuple returned contains the environment's next observation and reward for that action. The time_step_spec() and action_spec() methods in the environment return the specifications (types, shapes, bounds) of the time_step and action respectively.

print('Observation Spec:')
print('Action Spec:')
Observation Spec:
BoundedArraySpec(shape=(4,), dtype=dtype('float32'), name='observation', minimum=[-4.8000002e+00 -3.4028235e+38 -4.1887903e-01 -3.4028235e+38], maximum=[4.8000002e+00 3.4028235e+38 4.1887903e-01 3.4028235e+38])
Action Spec:
BoundedArraySpec(shape=(), dtype=dtype('int64'), name='action', minimum=0, maximum=1)

So, we see that observation is an array of 4 floats: the position and velocity of the cart, and the angular position and velocity of the pole. Since only two actions are possible (move left or move right), the action_spec is a scalar where 0 means "move left" and 1 means "move right."

time_step = env.reset()
print('Time step:')

action = np.array(1, dtype=np.int32)

next_time_step = env.step(action)
print('Next time step:')
Time step:
TimeStep(step_type=array(0, dtype=int32), reward=array(0., dtype=float32), discount=array(1., dtype=float32), observation=array([ 0.03506435, -0.04345531,  0.01258753,  0.00019599], dtype=float32))
Next time step:
TimeStep(step_type=array(1, dtype=int32), reward=array(1., dtype=float32), discount=array(1., dtype=float32), observation=array([ 0.03419525,  0.15148388,  0.01259145, -0.28848898], dtype=float32))

Usually we create two environments: one for training and one for evaluation. Most environments are written in pure python, but they can be easily converted to TensorFlow using the TFPyEnvironment wrapper. The original environment's API uses numpy arrays, the TFPyEnvironment converts these to/from Tensors for you to more easily interact with TensorFlow policies and agents.

train_py_env = suite_gym.load(env_name)
eval_py_env = suite_gym.load(env_name)

train_env = tf_py_environment.TFPyEnvironment(train_py_env)
eval_env = tf_py_environment.TFPyEnvironment(eval_py_env)


The algorithm that we use to solve an RL problem is represented as an Agent. In addition to the REINFORCE agent, TF-Agents provides standard implementations of a variety of Agents such as DQN, DDPG, TD3, PPO and SAC.

To create a REINFORCE Agent, we first need an Actor Network that can learn to predict the action given an observation from the environment.

We can easily create an Actor Network using the specs of the observations and actions. We can specify the layers in the network which, in this example, is the fc_layer_params argument set to a tuple of ints representing the sizes of each hidden layer (see the Hyperparameters section above).

actor_net = actor_distribution_network.ActorDistributionNetwork(

We also need an optimizer to train the network we just created, and a train_step_counter variable to keep track of how many times the network was updated.

optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)

train_step_counter = tf.compat.v2.Variable(0)

tf_agent = reinforce_agent.ReinforceAgent(


In TF-Agents, policies represent the standard notion of policies in RL: given a time_step produce an action or a distribution over actions. The main method is policy_step = policy.step(time_step) where policy_step is a named tuple PolicyStep(action, state, info). The policy_step.action is the action to be applied to the environment, state represents the state for stateful (RNN) policies and info may contain auxiliary information such as log probabilities of the actions.

Agents contain two policies: the main policy that is used for evaluation/deployment (agent.policy) and another policy that is used for data collection (agent.collect_policy).

eval_policy = tf_agent.policy
collect_policy = tf_agent.collect_policy

Metrics and Evaluation

The most common metric used to evaluate a policy is the average return. The return is the sum of rewards obtained while running a policy in an environment for an episode, and we usually average this over a few episodes. We can compute the average return metric as follows.

def compute_avg_return(environment, policy, num_episodes=10):

  total_return = 0.0
  for _ in range(num_episodes):

    time_step = environment.reset()
    episode_return = 0.0

    while not time_step.is_last():
      action_step = policy.action(time_step)
      time_step = environment.step(action_step.action)
      episode_return += time_step.reward
    total_return += episode_return

  avg_return = total_return / num_episodes
  return avg_return.numpy()[0]

# Please also see the metrics module for standard implementations of different
# metrics.

Replay Buffer

In order to keep track of the data collected from the environment, we will use the TFUniformReplayBuffer. This replay buffer is constructed using specs describing the tensors that are to be stored, which can be obtained from the agent using tf_agent.collect_data_spec.

replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(

For most agents, the collect_data_spec is a Trajectory named tuple containing the observation, action, reward etc.

Data Collection

As REINFORCE learns from whole episodes, we define a function to collect an episode using the given data collection policy and save the data (observations, actions, rewards etc.) as trajectories in the replay buffer.

def collect_episode(environment, policy, num_episodes):

  episode_counter = 0

  while episode_counter < num_episodes:
    time_step = environment.current_time_step()
    action_step = policy.action(time_step)
    next_time_step = environment.step(action_step.action)
    traj = trajectory.from_transition(time_step, action_step, next_time_step)

    # Add trajectory to the replay buffer

    if traj.is_boundary():
      episode_counter += 1

# This loop is so common in RL, that we provide standard implementations of
# these. For more details see the drivers module.

Training the agent

The training loop involves both collecting data from the environment and optimizing the agent's networks. Along the way, we will occasionally evaluate the agent's policy to see how we are doing.

The following will take ~3 minutes to run.


# (Optional) Optimize by wrapping some of the code in a graph using TF function.
tf_agent.train = common.function(tf_agent.train)

# Reset the train step

# Evaluate the agent's policy once before training.
avg_return = compute_avg_return(eval_env, tf_agent.policy, num_eval_episodes)
returns = [avg_return]

for _ in range(num_iterations):

  # Collect a few episodes using collect_policy and save to the replay buffer.
      train_env, tf_agent.collect_policy, collect_episodes_per_iteration)

  # Use data from the buffer and update the agent's network.
  experience = replay_buffer.gather_all()
  train_loss = tf_agent.train(experience)

  step = tf_agent.train_step_counter.numpy()

  if step % log_interval == 0:
    print('step = {0}: loss = {1}'.format(step, train_loss.loss))

  if step % eval_interval == 0:
    avg_return = compute_avg_return(eval_env, tf_agent.policy, num_eval_episodes)
    print('step = {0}: Average Return = {1}'.format(step, avg_return))
WARNING:tensorflow:From <ipython-input-16-235ae48023f9>:24: ReplayBuffer.gather_all (from tf_agents.replay_buffers.replay_buffer) is deprecated and will be removed in a future version.
Instructions for updating:
Use `as_dataset(..., single_deterministic_pass=True)` instead.
step = 25: loss = 0.14318394660949707
step = 50: loss = -0.5732654333114624
step = 50: Average Return = 61.70000076293945
step = 75: loss = 0.4367893934249878
step = 100: loss = -1.1474390029907227
step = 100: Average Return = 79.0999984741211
step = 125: loss = 0.8831806182861328
step = 150: loss = -0.050168514251708984
step = 150: Average Return = 184.10000610351562
step = 175: loss = 0.9362773895263672
step = 200: loss = -1.2963881492614746
step = 200: Average Return = 200.0
step = 225: loss = -1.6973439455032349
step = 250: loss = -3.887733221054077
step = 250: Average Return = 200.0



We can plot return vs global steps to see the performance of our agent. In Cartpole-v0, the environment gives a reward of +1 for every time step the pole stays up, and since the maximum number of steps is 200, the maximum possible return is also 200.

steps = range(0, num_iterations + 1, eval_interval)
plt.plot(steps, returns)
plt.ylabel('Average Return')
(5.855000400543213, 250.0)



It is helpful to visualize the performance of an agent by rendering the environment at each step. Before we do that, let us first create a function to embed videos in this colab.

def embed_mp4(filename):
  """Embeds an mp4 file in the notebook."""
  video = open(filename,'rb').read()
  b64 = base64.b64encode(video)
  tag = '''
  <video width="640" height="480" controls>
    <source src="data:video/mp4;base64,{0}" type="video/mp4">
  Your browser does not support the video tag.

  return IPython.display.HTML(tag)

The following code visualizes the agent's policy for a few episodes:

num_episodes = 3
video_filename = 'imageio.mp4'
with imageio.get_writer(video_filename, fps=60) as video:
  for _ in range(num_episodes):
    time_step = eval_env.reset()
    while not time_step.is_last():
      action_step = tf_agent.policy.action(time_step)
      time_step = eval_env.step(action_step.action)

WARNING:root:IMAGEIO FFMPEG_WRITER WARNING: input image is not divisible by macro_block_size=16, resizing from (400, 600) to (400, 608) to ensure video compatibility with most codecs and players. To prevent resizing, make your input image divisible by the macro_block_size or set the macro_block_size to None (risking incompatibility). You may also see a FFMPEG warning concerning speedloss due to data not being aligned.