CartPole DQN

Overview

This tutorial will show you how to solve the popular CartPole problem using deep Q-learning. The CartPole problem is as follows:

A pole is attached by an un-actuated joint to a cart, which moves along a frictionless track.
The system is controlled by applying a force of +1 or -1 to the cart.
The pendulum starts upright, and the goal is to prevent it from falling over.
A reward of +1 is provided for every timestep that the pole remains upright.
The episode ends when the pole is more than 15 degrees from vertical,
or the cart moves more than 2.4 units from the center.

Tutorial

This section will walk you through the steps of solving the CartPole problem with a deep Q-network. This tutorial is written for python 3.

Packages

You must first pip install the following packages: gym keras and numpy

DQN Agent

The first step of our implementation will be creating a DQNAgent object. This object will manage the state of our learning, and is independent of the CartPole problem. It has all the generic parts of a Q-learning agent and can be reused for other deep Q-learning applications. Every subsection will contain a part of the DQNAgent class you must implement.

Imports

Start by creating a file DQNAgent.py and include the following imports:

import numpy as np
from keras.layers import Input, Dense
from keras.optimizers import RMSprop
from keras.models import Model
from collections import deque

Constructor

The reason for each import will become apparent as our implementation continues. Next add a blank DQNAgent class with an empty constructor.

class DQNAgent:
    
    def __init__(self):
        pass

This class will take in all of our hyperparemeters, so let's update our constructor to take in those parameters. We also provide some default values for some of those hyperparameters.

class DQNAgent:
    
    def __init__(self, input_dim, output_dim, learning_rate=.005,
            mem_size=5000, batch_size=64, gamma=.99, decay_rate=.0002):
        pass

• input_dim is the number of input nodes for our DQN.
• output_dim is the number of output nodes for our DQN.
• learning_rate is a Keras parameter for our network describing how much we value new information.
• mem_size is the maximum number of instances allowed in our bucket for experience replay.
• batch_size is the number of experience tuples we train our model on each replay event.
• gamma is our discount factor for the Bellman equation update.
• decay_rate is the rate at which exploration probability decays.

Now for the next step, we complete our constructor by saving all of these parameters as instance variables, defining a neural network model, and defining a few other parameters.

def __init__(self, input_dim, output_dim, learning_rate=.005,
        mem_size=5000, batch_size=64, gamma=.99, decay_rate=.0002):

    # Save instance variables.
    self.input_dim = input_dim
    self.output_dim = output_dim
    self.batch_size = batch_size
    self.gamma = gamma
    self.decay_rate = decay_rate

    # Define other instance variables.
    self.explore_p = 1 # The current probability of taking a random action.
    self.memory = deque(maxlen=mem_size) # Define our experience replay bucket as a deque with size mem_size.

    # Define and compile our DQN. This network has 3 layers of 24 nodes. This is sufficient to solve
    # CartPole, but you should definitely tweak the architecture for other implementations.
    input_layer = Input(shape=(input_dim,))
    hl = Dense(24, activation="relu")(input_layer)
    hl = Dense(24, activation="relu")(hl)
    hl = Dense(24, activation="relu")(hl)
    output_layer = Dense(output_dim, activation="linear")(hl)
    self.model = Model(input_layer, output_layer)
    self.model.compile(loss="mse", optimizer=RMSprop(lr=learning_rate))

Act

The most fundamental part of a Q-learning problem is the ability for the agent to take an action. Actions are either determined by the current policy (based off Q-function values) or are picked randomly, depending on the current exploration probability. We now define an act function which, given the current state of the environment, determines which action to take next. Note that with OpenAI gym, actions correspond to integers (0, 1, 2, ...).

def act(self, state):
    # First, decay our explore probability
    self.explore_p *= 1 - self.decay_rate
    # With probability explore_p, randomly pick an action
    if self.explore_p > np.random.rand():
        return np.random.randint(self.output_dim)
    # Otherwise, find the action that should maximize future rewards according to our current Q-function policy.
    else:
        return np.argmax(self.model.predict(np.array([state]))[0])

Remember

One of the crucial parts of deep Q-learning is experience replay, where we store instances in a bucket and randomly draw from them to train our model. We now define the remember function, which stores the given experience tuple in that experience replay bucket for later sampling.

def remember(self, state, action, next_state, reward):
    # Create a blank state. Serves as next_state if this was the last experience tuple before the simulation ended.
    terminal_state = np.array([None]*self.input_dim) 
    # Add experience tuple to bucket. Bucket is a deque, so older tuple falls out on overflow.
    self.memory.append((state, action, terminal_state if next_state is None else next_state, reward))

Replay

The replay step is where experience tuples are randomly sampled from the bucket and are used to train the DQN. We now define the replay function to do just that.

def replay(self):

    # Only conduct a replay if we have enough experience to sample from.
    if len(self.memory) < self.batch_size:
        return

    # Pick random indices from the bucket without replacement. batch_size determines number of samples.
    idx = np.random.choice(len(self.memory), size=self.batch_size, replace=False)
    minibatch = np.array(self.memory)[idx]
    self.train(minibatch)
    
    # Extract the columns from our sample
    states = np.array(list(minibatch[:,0]))
    actions = minibatch[:,1]
    next_states = np.array(list(minibatch[:,2]))
    rewards = np.array(minibatch[:,3])

    # Compute a new estimate for each Q-value. This uses the second half of Bellman's equation.
    estimate = rewards + self.gamma * np.amax(self.model.predict(next_states), axis=1)

    # Get the network's current Q-value predictions for the states in this sample.
    predictions = self.model.predict(states)
    # Update the network's predictions with the new predictions we have.
    for i in range(len(predictions)):
        # Flag states as terminal (the last state before a simulation ended).
        terminal_state = (next_states[i] == np.array([None, None, None, None])).all()
        # Update each state's Q-value prediction with our new estimate.
        # Terminal states have no future, so set their Q-value to their immediate reward.
        predictions[i][actions[i]] = rewards[i] if terminal_state else estimate[i]

    # Propagate the new predictions through our network.
    self.model.fit(states, predictions, epochs=1, verbose=0)

This completes our DQNAgent object. Now let's move on to the actual driver of the learning process, which interacts with CartPole to drive learning.