DQN CartPole with Keras, based on https://github.com/udacity/deep-learning/blob/master/reinforcement/Q-learning-cart.ipynb

dqn_cartpole_keras.py

import gym
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam
from collections import deque

# Create the Cart-Pole game environment
env = gym.make('CartPole-v0')


class QNetwork:
    def __init__(self, learning_rate=0.01, state_size=4,
                 action_size=2, hidden_size=10):
        # state inputs to the Q-network
        self.model = Sequential()

        self.model.add(Dense(hidden_size, activation='relu',
                             input_dim=state_size))
        self.model.add(Dense(hidden_size, activation='relu'))
        self.model.add(Dense(action_size, activation='linear'))

        self.optimizer = Adam(lr=learning_rate)
        self.model.compile(loss='mse', optimizer=self.optimizer)


class Memory():
    def __init__(self, max_size=1000):
        self.buffer = deque(maxlen=max_size)

    def add(self, experience):
        self.buffer.append(experience)

    def sample(self, batch_size):
        idx = np.random.choice(np.arange(len(self.buffer)),
                               size=batch_size,
                               replace=False)
        return [self.buffer[ii] for ii in idx]


train_episodes = 1000          # max number of episodes to learn from
max_steps = 200                # max steps in an episode
gamma = 0.99                   # future reward discount

# Exploration parameters
explore_start = 1.0            # exploration probability at start
explore_stop = 0.01            # minimum exploration probability
decay_rate = 0.0001            # exponential decay rate for exploration prob

# Network parameters
hidden_size = 16               # number of units in each Q-network hidden layer
learning_rate = 0.001         # Q-network learning rate

# Memory parameters
memory_size = 10000            # memory capacity
batch_size = 32                # experience mini-batch size
pretrain_length = batch_size   # number experiences to pretrain the memory

mainQN = QNetwork(hidden_size=hidden_size, learning_rate=learning_rate)

###################################
## Populate the experience memory
###################################

# Initialize the simulation
env.reset()
# Take one random step to get the pole and cart moving
state, reward, done, _ = env.step(env.action_space.sample())
state = np.reshape(state, [1, 4])

memory = Memory(max_size=memory_size)

# Make a bunch of random actions and store the experiences
for ii in range(pretrain_length):
    # Uncomment the line below to watch the simulation
    # env.render()

    # Make a random action
    action = env.action_space.sample()
    next_state, reward, done, _ = env.step(action)
    next_state = np.reshape(next_state, [1, 4])

    if done:
        # The simulation fails so no next state
        next_state = np.zeros(state.shape)
        # Add experience to memory
        memory.add((state, action, reward, next_state))

        # Start new episode
        env.reset()
        # Take one random step to get the pole and cart moving
        state, reward, done, _ = env.step(env.action_space.sample())
        state = np.reshape(state, [1, 4])
    else:
        # Add experience to memory
        memory.add((state, action, reward, next_state))
        state = next_state

#############
## Training
#############
step = 0
for ep in range(1, train_episodes):
    total_reward = 0
    t = 0
    while t < max_steps:
        step += 1
        # Uncomment this next line to watch the training
        # env.render()

        # Explore or Exploit
        explore_p = explore_stop + (explore_start - explore_stop)*np.exp(-decay_rate*step)
        if explore_p > np.random.rand():
            # Make a random action
            action = env.action_space.sample()
        else:
            # Get action from Q-network
            Qs = mainQN.model.predict(state)[0]
            action = np.argmax(Qs)

        # Take action, get new state and reward
        next_state, reward, done, _ = env.step(action)
        next_state = np.reshape(next_state, [1, 4])
        total_reward += reward

        if done:
            # the episode ends so no next state
            next_state = np.zeros(state.shape)
            t = max_steps

            print('Episode: {}'.format(ep),
                  'Total reward: {}'.format(total_reward),
                  'Explore P: {:.4f}'.format(explore_p))

            # Add experience to memory
            memory.add((state, action, reward, next_state))

            # Start new episode
            env.reset()
            # Take one random step to get the pole and cart moving
            state, reward, done, _ = env.step(env.action_space.sample())
            state = np.reshape(state, [1, 4])
        else:
            # Add experience to memory
            memory.add((state, action, reward, next_state))
            state = next_state
            t += 1

        # Replay
        inputs = np.zeros((batch_size, 4))
        targets = np.zeros((batch_size, 2))

        minibatch = memory.sample(batch_size)
        for i, (state_b, action_b, reward_b, next_state_b) in enumerate(minibatch):
            inputs[i:i+1] = state_b
            target = reward_b
            if not (next_state_b == np.zeros(state_b.shape)).all(axis=1):
                target_Q = mainQN.model.predict(next_state_b)[0]
                target = reward_b + gamma * np.amax(mainQN.model.predict(next_state_b)[0])
            targets[i] = mainQN.model.predict(state_b)
            targets[i][action_b] = target
        mainQN.model.fit(inputs, targets, epochs=1, verbose=0)

Episode: 22 Total reward: 48.0 Explore P: 0.9598
Episode: 23 Total reward: 28.0 Explore P: 0.9571
Episode: 24 Total reward: 20.0 Explore P: 0.9552
Episode: 25 Total reward: 16.0 Explore P: 0.9537
Episode: 26 Total reward: 14.0 Explore P: 0.9524
Episode: 27 Total reward: 13.0 Explore P: 0.9512
Episode: 28 Total reward: 15.0 Explore P: 0.9497
Episode: 29 Total reward: 16.0 Explore P: 0.9482
Episode: 30 Total reward: 11.0 Explore P: 0.9472
Episode: 31 Total reward: 21.0 Explore P: 0.9452

Episode: 169 Total reward: 199.0 Explore P: 0.5193
Episode: 170 Total reward: 55.0 Explore P: 0.5165
Episode: 171 Total reward: 138.0 Explore P: 0.5095
Episode: 172 Total reward: 174.0 Explore P: 0.5009
Episode: 173 Total reward: 199.0 Explore P: 0.4913
Episode: 174 Total reward: 31.0 Explore P: 0.4898
Episode: 175 Total reward: 199.0 Explore P: 0.4803
Episode: 176 Total reward: 50.0 Explore P: 0.4780
Episode: 177 Total reward: 124.0 Explore P: 0.4722
Episode: 178 Total reward: 187.0 Explore P: 0.4636

Episode: 246 Total reward: 199.0 Explore P: 0.2822
Episode: 247 Total reward: 199.0 Explore P: 0.2769
Episode: 248 Total reward: 199.0 Explore P: 0.2716
Episode: 249 Total reward: 199.0 Explore P: 0.2665
Episode: 250 Total reward: 199.0 Explore P: 0.2614
Episode: 251 Total reward: 199.0 Explore P: 0.2565
Episode: 252 Total reward: 199.0 Explore P: 0.2516
Episode: 253 Total reward: 199.0 Explore P: 0.2468

https://gym.openai.com/evaluations/eval_DYySvAnSxuP3oHjZchxaQ

Could you tell me if it works robustly or just converges sometimes and what hyperparameters you used?
I tried to execute this code and it never learns anything and I don't know if it's because of the code or if I have some problem with my keras version.

('Episode: 580', 'Total reward: 12.0', 'Explore P: 0.4601')
('Episode: 581', 'Total reward: 13.0', 'Explore P: 0.4595')
('Episode: 582', 'Total reward: 10.0', 'Explore P: 0.4591')
('Episode: 583', 'Total reward: 8.0', 'Explore P: 0.4587')
('Episode: 584', 'Total reward: 10.0', 'Explore P: 0.4583')
('Episode: 585', 'Total reward: 8.0', 'Explore P: 0.4579')
('Episode: 586', 'Total reward: 9.0', 'Explore P: 0.4575')
('Episode: 587', 'Total reward: 15.0', 'Explore P: 0.4568')
('Episode: 588', 'Total reward: 9.0', 'Explore P: 0.4564')
('Episode: 589', 'Total reward: 9.0', 'Explore P: 0.4560')
('Episode: 590', 'Total reward: 8.0', 'Explore P: 0.4557')
('Episode: 591', 'Total reward: 13.0', 'Explore P: 0.4551')

Can you please tell me why you aren't using a different target network like it is mentioned in the dqn paper?