I've been trying to train a DDQN to play OpenAI Gym's CartPole-v1, but found that although it starts off well and starts getting full score (500) repeatedly (at around 600 episodes in the pic below), it then seems to go off the rails and do worse the more it plays.

Graph of getting good, then going really bad

I'm pretty new to ML so I'm not really sure what could cause this so I'm not sure how to start debugging (I've tried tweaking some of the hyper-parameters, but nothing seems to stop this trend).

If it helps, here's the (probably) relevant parts of my agent:

def _build_model(self):
    model = Sequential()
    model.add(Dense(24, input_dim=self.state_size, activation="relu"))
    model.add(Dense(24, activation="relu"))
    model.add(Dense(self.action_size, activation="linear"))
    model.compile(optimizer=Adam(lr=self.learning_rate), loss="mse")
    return model

def get_action(self, state):
    # Use random exploration for the current rate.
    if np.random.rand() < self.epsilon:
        return random.randrange(self.action_size)
    # Otherwise use the model to predict the rewards and select the max.
    q_values = self.model.predict(state)
    return np.argmax(q_values[0])

def replay(self, batch_size):
    if len(agent.memory) < minibatch_size:
    # Decay the exploration rate.
    self.epsilon *= self.epsilon_decay
    self.epsilon = max(self.epsilon_min, self.epsilon)
    minibatch = random.sample(self.memory, minibatch_size)
    state_batch, q_values_batch = [], []
    for state, action, reward, next_state, done in minibatch:
        # Get predictions for all actions for the current state.
        q_values = self.model.predict(state)
        # If we're not done, add on the future predicted reward at the discounted rate.
        if done:
            q_values[0][action] = reward
            f = self.target_model.predict(next_state)
            future_reward = max(self.target_model.predict(next_state)[0])
            q_values[0][action] = reward + self.gamma * future_reward
    # Re-fit the model to move it closer to this newly calculated reward.
    self.model.fit(np.array(state_batch), np.array(q_values_batch), batch_size=batch_size, epochs=1, verbose=0)
def update_weights(self):
    weights = self.model.get_weights()
    target_weights = self.target_model.get_weights()
    for i in range(len(target_weights)):
        target_weights[i] = weights[i] * self.tau + target_weights[i] * (1 - self.tau)

And the full notebook is here.


1 Answer 1


This is called "catastrophic forgetting" and can be a serious problem in many RL scenarios.

If you trained a neural network to recognise cats and dogs and did the following:

  • Train it for many epochs on a full dataset until you got a high accuracy.

  • Continue to train it, but remove all the cat pictures.

Then in a relatively short space of time, the NN would start to lose accuracy. It would forget what a cat looks like. It would learn that its task was to switch the dog prediction as high as possible, just because on average everything in the training population was a dog.

Something very similar happens in your DQN experience replay memory. Once it gets good at a task, it may only experience success. Eventually, only successful examples are in its memory. The NN forgets what failure looks like (what the states are, and what it should predict for their values), and predicts high values for everything.

When something bad happens and the NNs high predicted value is completely wrong, the error can be high, and the NN may have incorrectly "linked" its state representations so that it cannot distinguish which parts of the feature space are the cause of this. This creates odd effects in terms of what it learns about values of all states. Often the NN will behave incorrectly for a few episodes but then re-learn optimal behaviour. But it is also possible that it completely breaks and never recovers.

There is lots of active research into catastrophic forgetting and I suggest you search that term to find out some of the many types of mitigation you could use.

For Cartpole, I found a very simple hack made the learning very stable. Simply keep aside some percentage of replay memory stocked with the initial poor performing random exploration. Reserving say 10% to this long term memory is enough to make learning in Cartpole rock solid, as the NN always has a few examples of what not to do. The idea unfortunately does not scale well to more complex environments, but it is a nice demonstration. For a more sophisticated look at similar solutions you could see the paper "The importance of experience replay database composition in deep reinforcement learning"

  • $\begingroup$ Thanks for the excellent description! That hadn't been mentioned in the videos/tutorials I'd read but it seems to make sense. I'll give your idea a shot, and do some more digging into this problem. Thanks! $\endgroup$ Commented Jul 20, 2019 at 9:35
  • 2
    $\begingroup$ FYI this approach of keeping some of the initial experiences in the replay buffer is explored in "The importance of experience replay database composition in deep reinforcement learning". This paper, as a whole, provides a good answer to this question. Another relevant paper is "Implementing the Deep Q-Network". $\endgroup$ Commented Jul 20, 2019 at 10:11
  • $\begingroup$ @MasterScrat: thank you, I have added a link to the paper in the answer $\endgroup$ Commented Jul 20, 2019 at 10:28
  • $\begingroup$ Very well explained, the cat/dog analogy is quite helpful! $\endgroup$ Commented Jul 30, 2022 at 0:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.