I'm trying to gain an intuitive understanding of deep reinforcement learning. In deep Q-networks (DQN) we store all actions/environments/rewards in a memory array and at the end of the episode, "replay" them through our neural network. This makes sense because we are trying to build out our rewards matrix and see if our episode ended in reward, scale that back through our matrix.

I would think the sequence of actions that led to the reward state is what is important to capture - this sequence of actions (and not the actions independently) are what led us to our reward state.

In the Atari-DQN paper by Mnih and many tutorials since we see the practice of random sampling from the memory array and training. So if we have a memory of:

$(action\,a, state\,1) \rightarrow (action\,b, state\,2) \rightarrow (action\,c, state\,3) \rightarrow (action\,d, state\,4) \rightarrow reward!$

We may train a mini-batch of:

[(action c state 3), (action b, state 2), reward!]

The reason given is:

Second, learning directly from consecutive samples is inefficient, due to the strong correlations between the samples; randomizing the samples breaks these correlations and therefore reduces the variance of the updates.

or from this pytorch tutorial:

By sampling from it randomly, the transitions that build up a batch are decorrelated. It has been shown that this greatly stabilizes and improves the DQN training procedure.

My intuition would tell me the sequence is what is most important in reinforcement learning. Most episodes have a delayed reward so most action/states do not have a reward (and are not "reinforced"). The only way to bring a portion of the reward to these previous states is to retroactively break the reward out across the sequence (through the future_reward in the Q algorithm of reward + reward * learning_rate(future_reward))

A random sampling of the memory bank breaks our sequence, how does that help when you are trying to back-fill a Q (reward) matrix?

Perhaps this is more similar to a Markov model where every state should be considered independent? Where is the error in my intuition?


The de-correlation effect is more important than following sequence of trajectories in this case.

Single step Q-learning does not rely on trajectories to learn. It is slightly less efficient to do this in TD learning - a Q($\lambda$) algorithm which averages over multiple trajectory lengths would maybe work better if it were not for the instability of using function approximators.

Instead, the DQN-based learning bootstraps across single steps (State, Action, Reward, Next State). It doesn't need longer trajectories. And in fact due to bias caused by correlation, the neural network might suffer for it if you tried. Even with experience replay, the bootstrapping - using one set of estimates to refine another - can be unstable. So other stabilising influences are beneficial too, such as using a frozen copy of the network to estimate the TD target $R + \text{max}_{a'} Q(S', a')$ - sometimes written $R + \text{max}_{a'} \hat{q}(S', a', \theta^{\bar{ }})$ where $\theta$ are the learnable parameters for $\hat{q}$ function.

It might still be possible to use longer trajectories, sampled randomly, to get a TD target estimate based on more steps. This can be beneficial for reducing bias from bootstrapping, at the expense of adding variance due to sampling from larger space of possible trajectories (and "losing" parts of trajectories or altering predicted reward because of exploratory actions). However, the single-step method presented by DQN has shown success, and it is not clear which problems would benefit from longer trajectories. You might like to experiment with options though . . . it is not an open-and-shut case, and since the DQN paper, various other refinements have been published.

| improve this answer | |
  • $\begingroup$ Great answer - thank you. How would this work in cases where the environment has a state where, depending on history, the expected reward is different? For instance - say we are running a driving simulator and you are approaching an intersection. There is a car crossing the intersection directly in front of you. Whether you swerve left or right (and by how much) is completely dependent on the states beforehand. The current state is just a car in front of yours - and sometimes you would turn left in this scenario and others, right. If all we have is a single prior state, what if that isnt enoug $\endgroup$ – ZAR Nov 19 '17 at 21:13
  • $\begingroup$ @ZAR: If reward has other non-random dependencies than your state, then you have defined state incorrectly (i.e. in a way that breaks the Markov property that RL theory relies upon). In your example, the state should include belief states about positions of other cars, not just simply what the agent sees on a specific time step. There are also POMDPs, which might be a better model if older observations are less strongly linked to reward distribution. $\endgroup$ – Neil Slater Nov 19 '17 at 21:20
  • 1
    $\begingroup$ The DQN is usually trained with around 5 sequential images. So your state is not just the image of the screen at the current moment. Another implementation of state representation is the difference of two sequential images (which capture changes e.g. in speed/direction). LSTMs (instead of the feedforward NNs) can help you deal with non-Markovian tasks. $\endgroup$ – Constantinos Nov 20 '17 at 7:12
  • 2
    $\begingroup$ @ZAR: You are confusing observation with state. In an MDP, the state is defined to have the Markov property. Which means you often have to go beyond single observations to build a good representation of state. In the case of Atari (including Pong), the authors chose multiple frames as their state, which does then include recent movement. $\endgroup$ – Neil Slater Nov 20 '17 at 7:53
  • 1
    $\begingroup$ Imagine this as the fog of war in a strategy game. You have no clue about what your opponent is doing but you can have partial information (e.g. about his progress) by inspecting its units (e.g. upgrades and armory of the units). From this you can have an estimate of the current state of the game and try to act optimally given your beliefs which are based on your observations. $\endgroup$ – Constantinos Nov 20 '17 at 16:00

The DQN uses experience replay to break correlations between sequential experiences. It is viewed that for every state, the next state is going to be affected by the current action, therefore, taking experiences sequentially would result in instabilities due to internal correlations between experiences. An experience consists of a state, an action, a reward and the next state; all that is needed to learn, to the very least in temporal difference fashion. As such, experience replay enables us to combine Monte Carlo (which uses full episodes) and temporal difference (which uses single experiences) in one but more robust scheme. You realize that what is needed is not the size of a single reward but an indication as to whether the agent is on the right track, which is best given by an average of several experiences.

Also, using same network parameters to get both the prediction and the target Q-values is like updating a guess with another guess, which is like "a dog chasing its own tail." The use of a target network which is just a clone (stale copy) of the prediction network would help break this other kind of correlation between targets and prediction. While the prediction network is updated using the experiences, the target network parameters are periodically updated, say after every so many updates of the prediction network, its weights are copied to the target network.

| improve this answer | |
  • $\begingroup$ So does that mean the target network works on old weights? What does that mean for the weights? $\endgroup$ – mikanim Jun 6 at 22:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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