How to use a custom metric with Tensorflow Agents
TensorFlow Reinforcement learning

How to use a custom metric with Tensorflow Agents

How to define a new Tensorflow Agents metric and add it to the driver
Bartosz Mikulski
Bartosz Mikulski – 4 min read

In this article, I am going to implement a custom Tensorflow Agents metric that calculates the maximal discounted reward.

First, I have to import the metric-related modules and the driver module (the driver runs the simulation). Additionally, I need an environment. I’m going to use the one I implemented in this article.

from tf_agents.metrics import tf_py_metric
from tf_agents.metrics import py_metric
from tf_agents.drivers import py_driver
from tf_agents.drivers import dynamic_episode_driver

My metric needs to store the rewards and discounts from the current episode and the maximal discounted total score. For that, I need two arrays (for the episode scores) and one variable to keep the maximal reward.

class MaxEpisodeScoreMetric(py_metric.PyStepMetric):
  def __init__(self, name='MaxEpisodeScoreMetric'):
    super(py_metric.PyStepMetric, self).__init__(name)
    self.rewards = []
    self.discounts = []
    self.max_discounted_reward = None
    self.reset()

The reset function is mandatory, and it allows the metric instance to be reused by separate driver runs.

#add it inside the MaxEpisodeScoreMetric class
  def reset(self):
    self.rewards = []
    self.discounts = []
    self.max_discounted_reward = None

In the call function, I am going to copy the reward and discount of the current step to the arrays. Then, if the current step is also the last step of an episode, I am going to calculate the discounted reward using the Bellman equation.

After that, I compare the total discounted reward of the current episode with the maximal reward. If I got a value larger than the current maximum, I would replace the maximum with the new value.

Because the instance is not reset between episodes, I need to clear the lists I use to keep the episode rewards and discounts.

#add it inside the MaxEpisodeScoreMetric class
def call(self, trajectory):
    self.rewards += trajectory.reward
    self.discounts += trajectory.discount

    if(trajectory.is_last()):
      adjusted_discounts = [1.0] + self.discounts # because a step has its value + the discount of the NEXT step (Bellman equation)
      adjusted_discounts = adjusted_discounts[:-1] # dropping the discount of the last step because it is not followed by a next step, so the value is useless
      discounted_reward = np.sum(np.multiply(self.rewards, adjusted_discounts))
      print(self.rewards, adjusted_discounts, discounted_reward)

      if self.max_discounted_reward == None:
        self.max_discounted_reward = discounted_reward

      if discounted_reward > self.max_discounted_reward:
        self.max_discounted_reward = discounted_reward

      self.rewards = []
      self.discounts = []

In the result function, I don’t need to perform any additional operations, so I return the maximal discounted total reward.

#add it inside the MaxEpisodeScoreMetric class
  def result(self):
    return self.max_discounted_reward

I want to use my metric as a Tensorflow metric, so I had to wrap it with a class extending TFPyMetric.

class TFMaxEpisodeScoreMetric(tf_py_metric.TFPyMetric):

  def __init__(self, name='MaxEpisodeScoreMetric', dtype=tf.float32):
    py_metric = MaxEpisodeScoreMetric()

    super(TFMaxEpisodeScoreMetric, self).__init__(
        py_metric=py_metric, name=name, dtype=dtype)

Finally, I can add the metric to the driver’s observers and run the driver.

#tf_env is from the article mentioned in the second paragraph
tf_policy = random_tf_policy.RandomTFPolicy(action_spec=tf_env.action_spec(),
                                            time_step_spec=tf_env.time_step_spec())

max_score = TFMaxEpisodeScoreMetric()

observers = [max_score]
driver = dynamic_episode_driver.DynamicEpisodeDriver(tf_env, tf_policy, observers, num_episodes=1000)

final_time_step, policy_state = driver.run()

print('Max score:', max_score.result().numpy())

Result:

Max score: 1.715
Older post

How to use a behavior policy with Tensorflow Agents

Random and scripted behavior policies

Newer post

How to train a Reinforcement Learning Agent using Tensorflow Agents

The reinforcement learning loop with Tensorflow Agents

Related Posts

Reinforcement Learning TensorFlow

Using Boltzmann distribution as the exploration policy in TensorFlow-agent reinforcement learning models

There is a whole spectrum of exploration strategies between random and greedy policies.
Published on: 04 Nov 2019

Data Science TensorFlow

How to turn Pandas data frame into time-series input for RNN

From Pandas dataframe to RNN input
Published on: 14 Aug 2019

TensorFlow Deep learning

How to automatically select the hyperparameters of a ResNet neural network

Training ResNet network for multiclass image classification using keras-tuner
Published on: 12 Aug 2019