Hands-On Q-Learning with Python

Building a DQN class

We're going to put together a basic DQN algorithm that actually progresses through the environment and keeps the CartPole in motion for more than a few steps. We'll build on the performance of this basic algorithm after we see its results, but this will be sufficient to get a solution up and running.

We start by initializing the hyperparameters and the action space:

class DQN:    def __init__(self):        self.epsilon = epsilon        self.gamma = gamma        self.alpha = alpha                self.action_space = env.action_space.n

We're now creating a three-layer sequential network in Keras using relu activation functions for the input and hidden layers:

        self.model = Sequential()        self.model.add.(Dense(24, input_shape=(observation_space,), activation=