model.add(Dense(nb_neuron_output, activation='softmax'))
model.summary()
# Create the DQN agent with the memory and policy hyper-parameters
memory = SequentialMemory(limit=50000, window_length=1)
policy = BoltzmannQPolicy()
dqn = DQNAgent(model=model, nb_actions=nb_neuron_output, memory=memory, nb_steps_warmup=10,
target_model_update=1e-2, policy=policy)
dqn.compile(Adam(lr=1e-3), metrics=['mae', 'acc'])
# Create the metrics and pass it to the model as a callback for training
metrics = Metrics(dqn, env)
dqn.fit(env, nb_steps=1000000, visualize=False, verbose=2, callbacks=[metrics])
# the filename has the form modelName_memoryType|memorySize_policy_numberOfSteps
fileName = '1D__Sequential50000_BoltzmannQ_1000000steps(0)'
# Write the metrics into a file for post treatment
f1=open('./output/' + fileName + '.txt', 'w+')
f1.write(metrics.export_to_text())
f1.close()
# Optionally export figures
metrics.export_figs(fileName)
# Save network weights for later tests
dqn.save_weights('./output/' + fileName + '.h5f', overwrite=True)
# Test once after the training
dqn.test(env, nb_episodes=1, visualize=True)
