Merge pull request #2306 from peastman/tutorials

%% Cell type:markdown id: tags:

# Tutorial Part 18: Using Reinforcement Learning to Play Pong

This notebook demonstrates using reinforcement learning to train an agent to play Pong.

The first step is to create an `Environment` that implements this task.  Fortunately,

OpenAI Gym already provides an implementation of Pong (and many other tasks appropriate

for reinforcement learning).  DeepChem's `GymEnvironment` class provides an easy way to

use environments from OpenAI Gym.  We could just use it directly, but in this case we

subclass it and preprocess the screen image a little bit to make learning easier.

## Colab

This tutorial and the rest in this sequence are designed to be done in Google colab. If you'd like to open this notebook in colab, you can use the following link.

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/deepchem/deepchem/blob/master/examples/tutorials/18_Using_Reinforcement_Learning_to_Play_Pong.ipynb)

## Setup

To run DeepChem within Colab, you'll need to run the following cell of installation commands. This will take about 5 minutes to run to completion and install your environment. To install `gym` you should also use `pip install 'gym[atari]'` (We need the extra modifier since we'll be using an atari game). We'll add this command onto our usual Colab installation commands for you

%% Cell type:code id: tags:

``` python

!curl -Lo conda_installer.py https://raw.githubusercontent.com/deepchem/deepchem/master/scripts/colab_install.py

import conda_installer

conda_installer.install()

!/root/miniconda/bin/conda info -e

```

%% Output

      % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

                                     Dload  Upload   Total   Spent    Left  Speed

    
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
100  3489  100  3489    0     0  89461      0 --:--:-- --:--:-- --:--:-- 91815

    add /root/miniconda/lib/python3.6/site-packages to PYTHONPATH

    all packages is already installed

    # conda environments:

    #

    base                  *  /root/miniconda

%% Cell type:code id: tags:

``` python

!pip install --pre deepchem

import deepchem

deepchem.__version__

```

%% Output

    Requirement already satisfied: deepchem in /usr/local/lib/python3.6/dist-packages (2.4.0rc1.dev20200805145259)

    Requirement already satisfied: joblib in /usr/local/lib/python3.6/dist-packages (from deepchem) (0.16.0)

    Requirement already satisfied: scipy in /usr/local/lib/python3.6/dist-packages (from deepchem) (1.4.1)

    Requirement already satisfied: numpy in /usr/local/lib/python3.6/dist-packages (from deepchem) (1.18.5)

    Requirement already satisfied: pandas in /usr/local/lib/python3.6/dist-packages (from deepchem) (1.0.5)

    Requirement already satisfied: scikit-learn in /usr/local/lib/python3.6/dist-packages (from deepchem) (0.22.2.post1)

    Requirement already satisfied: python-dateutil>=2.6.1 in /usr/local/lib/python3.6/dist-packages (from pandas->deepchem) (2.8.1)

    Requirement already satisfied: pytz>=2017.2 in /usr/local/lib/python3.6/dist-packages (from pandas->deepchem) (2018.9)

    Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.6/dist-packages (from python-dateutil>=2.6.1->pandas->deepchem) (1.15.0)

    '2.4.0-rc1.dev'

%% Cell type:code id: tags:

``` python

!pip install 'gym[atari]'

```

%% Output

    Requirement already satisfied: gym[atari] in /usr/local/lib/python3.6/dist-packages (0.17.2)

    Requirement already satisfied: cloudpickle<1.4.0,>=1.2.0 in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (1.3.0)

    Requirement already satisfied: scipy in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (1.4.1)

    Requirement already satisfied: pyglet<=1.5.0,>=1.4.0 in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (1.5.0)

    Requirement already satisfied: numpy>=1.10.4 in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (1.18.5)

    Requirement already satisfied: Pillow; extra == "atari" in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (7.0.0)

    Requirement already satisfied: opencv-python; extra == "atari" in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (4.1.2.30)

    Requirement already satisfied: atari-py~=0.2.0; extra == "atari" in /usr/local/lib/python3.6/dist-packages (from gym[atari]) (0.2.6)

    Requirement already satisfied: future in /usr/local/lib/python3.6/dist-packages (from pyglet<=1.5.0,>=1.4.0->gym[atari]) (0.16.0)

    Requirement already satisfied: six in /usr/local/lib/python3.6/dist-packages (from atari-py~=0.2.0; extra == "atari"->gym[atari]) (1.15.0)

%% Cell type:code id: tags:

``` python

import deepchem as dc

import numpy as np

class PongEnv(dc.rl.GymEnvironment):

  def __init__(self):

    super(PongEnv, self).__init__('Pong-v0')

    self._state_shape = (80, 80)

  @property

  def state(self):

    # Crop everything outside the play area, reduce the image size,

    # and convert it to black and white.

    cropped = np.array(self._state)[34:194, :, :]

    reduced = cropped[0:-1:2, 0:-1:2]

    grayscale = np.sum(reduced, axis=2)

    bw = np.zeros(grayscale.shape)

    bw[grayscale != 233] = 1

    return bw

  def __deepcopy__(self, memo):

    return PongEnv()

env = PongEnv()

```

%% Cell type:markdown id: tags:

Next we create a network to implement the policy.  We begin with two convolutional layers to process

the image.  That is followed by a dense (fully connected) layer to provide plenty of capacity for game

logic.  We also add a small Gated Recurrent Unit.  That gives the network a little bit of memory, so

it can keep track of which way the ball is moving.

We concatenate the dense and GRU outputs together, and use them as inputs to two final layers that serve as the

network's outputs.  One computes the action probabilities, and the other computes an estimate of the

state value function.

We also provide an input for the initial state of the GRU, and returned its final state at the end.  This is required by the learning algorithm

%% Cell type:code id: tags:

``` python

import tensorflow as tf

from tensorflow.keras.layers import Input, Concatenate, Conv2D, Dense, Flatten, GRU, Reshape

class PongPolicy(dc.rl.Policy):

    def __init__(self):

        super(PongPolicy, self).__init__(['action_prob', 'value', 'rnn_state'], [np.zeros(16)])

    def create_model(self, **kwargs):

        state = Input(shape=(80, 80))

        rnn_state = Input(shape=(16,))

        conv1 = Conv2D(16, kernel_size=8, strides=4, activation=tf.nn.relu)(Reshape((80, 80, 1))(state))

        conv2 = Conv2D(32, kernel_size=4, strides=2, activation=tf.nn.relu)(conv1)

        dense = Dense(256, activation=tf.nn.relu)(Flatten()(conv2))

        gru, rnn_final_state = GRU(16, return_state=True, return_sequences=True)(

            Reshape((-1, 256))(dense), initial_state=rnn_state)

        concat = Concatenate()([dense, Reshape((16,))(gru)])

        action_prob = Dense(env.n_actions, activation=tf.nn.softmax)(concat)

        value = Dense(1)(concat)

        return tf.keras.Model(inputs=[state, rnn_state], outputs=[action_prob, value, rnn_final_state])

policy = PongPolicy()

```

%% Cell type:markdown id: tags:

We will optimize the policy using the Asynchronous Advantage Actor Critic (A3C) algorithm.  There are lots of hyperparameters we could specify at this point, but the default values for most of them work well on this problem.  The only one we need to customize is the learning rate.

%% Cell type:code id: tags:

``` python

# from deepchem.models.optimizers import Adam

# a3c = dc.rl.A3C(env, policy, model_dir='model', optimizer=Adam(learning_rate=0.0002))

```

%% Cell type:markdown id: tags:

Optimize for as long as you have patience to.  By 1 million steps you should see clear signs of learning.  Around 3 million steps it should start to occasionally beat the game's built in AI.  By 7 million steps it should be winning almost every time.  Running on my laptop, training takes about 20 minutes for every million steps.

%% Cell type:code id: tags:

``` python

# # Change this to train as many steps as you have patience for.

# a3c.fit(1000)

```

%% Cell type:markdown id: tags:

Let's watch it play and see how it does!

%% Cell type:code id: tags:

``` python

# # This code doesn't work well on Colab

# env.reset()

# while not env.terminated:

#     env.env.render()

#     env.step(a3c.select_action(env.state))

```

%% Cell type:markdown id: tags:

# Congratulations! Time to join the Community!

Congratulations on completing this tutorial notebook! If you enjoyed working through the tutorial, and want to continue working with DeepChem, we encourage you to finish the rest of the tutorials in this series. You can also help the DeepChem community in the following ways:

## Star DeepChem on [GitHub](https://github.com/deepchem/deepchem)

This helps build awareness of the DeepChem project and the tools for open source drug discovery that we're trying to build.

## Join the DeepChem Gitter

The DeepChem [Gitter](https://gitter.im/deepchem/Lobby) hosts a number of scientists, developers, and enthusiasts interested in deep learning for the life sciences. Join the conversation!