I am trying to train a DQN Agent to solve AI Gym's Cartpole-v0 environment. I have started with this person's implementation just to get some hands-on experience. What I noticed is that during training, after many episodes the agent finds the solution and is able to keep the pole upright for the maximum amount of timesteps. However, after further training, the policy looks like it becomes more stochastic and it can't keep the pole upright anymore and goes in and out of a good policy. I'm pretty confused by this why wouldn't further training and experience help the agent? At episodes my epsilon for random action becomes very low, so it should be operating on just making the next prediction. So why does it on some training episodes fail to keep the pole upright and on others it succeeds?
Here is a picture of my reward-episode curve during the training process of the above linked implementation.
This actually looks fairly normal to me, in fact I guessed your results were from CartPole before reading the whole question.
I have a few suggestions:
When you're plotting results, you should plot averages over a few random seeds. Not only is this generally good practice (it shows how sensitive your algo is to seeds), it'll smooth out your graphs and give you a better understanding of the "skill" of your agent. Don't forget, the environment and the policy are stochastic, so it's not completely crazy that your agent exhibits this type of behavior.
Assuming you're implementing e-greedy exploration, what's your epsilon value? Are you decaying it over time? The issue could also be that your agent is still exploring a lot even after it found a good policy.
Have you played around with hyperparameters, like learning rate, epsilon, network size, replay buffer size, etc? Those can also be the culprit.
Related
I finished going over the WGAN paper: WGAN Paper Link
After reading the algorithm provided by the writers I find it odd that they would refer to the network as an adversarial network.
In the first part of the algorithm a 'critic' is trained to optimality and they show this critic approximates the Wasserstein distance between our generator distribution and the real distribution. We then take this approximation and update the parameters of the generator distribution in the direction of the gradient of the critic. So in a sense we're just approximating a loss function and then we tell the generator in what direction is best to go. so a critic is a very good name for this, but calling it an adversarial network implies that the generator and the critic are at odds. Any ideas why this should still be nicknamed an adversarial network?
The name "adversarial" does not come from this paper, it comes from the GAN itself, this paper is merely an incremental work on top (and thus is not renaming anything). The reason why the "original" GAN is called Generative Adversarial Network is because it is trained in a form of a two-player, competitive game, where a generator task is to fool discriminator, and discriminators task is to well, not be fooled. This is the "at odds" part. And it is indeed critical to the whole system, vast majority of problems of GANs, that spawned hundreds of papers (like the one above) comes from the fact that greedy optimization of 2 player games has much more chaotic dynamics, and will not "just converge with small enough learning rate" that normal minimization of (smooth enough) loss function would. From math perspective, the subtle difference that makes things chaotic is that gradients that train discriminator are not back-propagated to the generator. Otherwise generator would be "helping" discriminator. Because of this stop gradient the emerging dynamics is no longer a gradient vector field of any loss, and instead it is a dynamical system emerging from simultaneous minimization of 2 functions (also called 2 player games).
I have a concern in understanding why a target network is necessary in DQN? I’m reading paper on “human-level control through deep reinforcement learning”
I understand Q-learning. Q-learning is value-based reinforcement learning algorithm that learns “optimal” probability distribution between state-action that will maximize it’s long term discounted reward over a sequence of timesteps.
The Q-learning is updated using the bellman equation, and a single step of the q-learning update is given by
Q(S, A) = Q(S, A) + $\alpha$[R_(t+1) + $\gamma$ (Q(s’,a;’) - Q(s,a)]
Where alpha and gamma are learning and discount factors.
I can understand that the reinforcement learning algorithm will become unstable and diverge.
The experience replay buffer is used so that we do not forget past experiences and to de-correlate datasets provided to learn the probability distribution.
This is where I fail.
Let me break the paragraph from the paper down here for discussion
The fact that small updates to $Q$ may significantly change the policy and therefore change the data distribution — understood this part. Changes to Q-network periodically may lead to unstability and changes in distribution. For example, if we always take a left turn or something like this.
and the correlations between the action-values (Q) and the target values r + $gamma$ (argmax(Q(s’,a’)) — This says that the reward + gamma * my prediction of the return given that I take what I think is the best action in the current state and follow my policy from then on.
We used an iterative update that adjusts the action-values (Q) towards target values that are only periodically updated, thereby reducing correlations with the target.
So, in summary a target network required because the network keeps changing at each timestep and the “target values” are being updated at each timestep?
But I do not understand how it is going to solve it?
So, in summary a target network required because the network keeps changing at each timestep and the “target values” are being updated at each timestep?
The difference between Q-learning and DQN is that you have replaced an exact value function with a function approximator. With Q-learning you are updating exactly one state/action value at each timestep, whereas with DQN you are updating many, which you understand. The problem this causes is that you can affect the action values for the very next state you will be in instead of guaranteeing them to be stable as they are in Q-learning.
This happens basically all the time with DQN when using a standard deep network (bunch of layers of the same size fully connected). The effect you typically see with this is referred to as "catastrophic forgetting" and it can be quite spectacular. If you are doing something like moon lander with this sort of network (the simple one, not the pixel one) and track the rolling average score over the last 100 games or so, you will likely see a nice curve up in score, then all of a sudden it completely craps out starts making awful decisions again even as your alpha gets small. This cycle will continue endlessly regardless of how long you let it run.
Using a stable target network as your error measure is one way of combating this effect. Conceptually it's like saying, "I have an idea of how to play this well, I'm going to try it out for a bit until I find something better" as opposed to saying "I'm going to retrain myself how to play this entire game after every move". By giving your network more time to consider many actions that have taken place recently instead of updating all the time, it hopefully finds a more robust model before you start using it to make actions.
On a side note, DQN is essentially obsolete at this point, but the themes from that paper were the fuse leading up to the RL explosion of the last few years.
Call for experts in deep learning.
Hey, I am recently working on training images using tensorflow in python for tone mapping. To get the better result, I focused on using perceptual loss introduced from this paper by Justin Johnson.
In my implementation, I made the use of all 3 parts of loss: a feature loss that extracted from vgg16; a L2 pixel-level loss from the transferred image and the ground true image; and the total variation loss. I summed them up as the loss for back propagation.
From the function
yˆ=argminλcloss_content(y,yc)+λsloss_style(y,ys)+λTVloss_TV(y)
in the paper, we can see that there are 3 weights of the losses, the λ's, to balance them. The value of three λs are probably fixed throughout the training.
My question is that does it make sense if I dynamically change the λ's in every epoch(or several epochs) to adjust the importance of these losses?
For instance, the perceptual loss converges drastically in the first several epochs yet the pixel-level l2 loss converges fairly slow. So maybe the weight λs should be higher for the content loss, let's say 0.9, but lower for others. As the time passes, the pixel-level loss will be increasingly important to smooth up the image and to minimize the artifacts. So it might be better to adjust it higher a bit. Just like changing the learning rate according to the different epochs.
The postdoc supervises me straightly opposes my idea. He thought it is dynamically changing the training model and could cause the inconsistency of the training.
So, pro and cons, I need some ideas...
Thanks!
It's hard to answer this without knowing more about the data you're using, but in short, dynamic loss should not really have that much effect and may have opposite effect altogether.
If you are using Keras, you could simply run a hyperparameter tuner similar to the following in order to see if there is any effect (change the loss accordingly):
https://towardsdatascience.com/hyperparameter-optimization-with-keras-b82e6364ca53
I've only done this on smaller models (way too time consuming) but in essence, it's best to keep it constant and also avoid angering off your supervisor too :D
If you are running a different ML or DL library, there are optimizer for each, just Google them. It may be best to run these on a cluster and overnight, but they usually give you a good enough optimized version of your model.
Hope that helps and good luck!
I have implemented a custom openai gym environment for a game similar to http://curvefever.io/, but with discreet actions instead of continuous. So my agent can in each step go in one of four directions, left/up/right/down. However one of these actions will always lead to the agent crashing into itself, since it cant "reverse".
Currently I just let the agent take any move, and just let it die if it makes an invalid move, hoping that it will eventually learn to not take that action in that state. I have however read that one can set the probabilities for making an illegal move zero, and then sample an action. Is there any other way to tackle this problem?
You can try to solve this by 2 changes:
1: give current direction as an input and give reward of maybe +0.1 if it takes the move which does not make it crash, and give -0.7 if it make a backward move which directly make it crash.
2: If you are using neural network and Softmax function as activation function of last layer, multiply all outputs of neural network with a positive integer ( confidence ) before giving it to Softmax function. it can be in range of 0 to 100 as i have experience more than 100 will not affect much. more the integer is the more confidence the agent will have to take action for a given state.
If you are not using neural network or say, deep learning, I suggest you to learn concepts of deep learning as your environment of game seems complex and a neural network will give best results.
Note: It will take huge amount of time. so you have to wait enough to train the algorithm. i suggest you not to hurry and let it train. and i played the game, its really interesting :) my wishes to make AI for the game :)
I am busy coding reinforcement learning agents for the game Pac-Man and came across Berkeley's CS course's Pac-Man Projects, specifically the reinforcement learning section.
For the approximate Q-learning agent, feature approximation is used. A simple extractor is implemented in this code. What I am curious about is why, before the features are returned, they are scaled down by 10? By running the solution without the factor of 10 you can notice that Pac-Man does significantly worse, but why?
After running multiple tests it turns out that the optimal Q-value can diverge wildly away. In fact, the features can all become negative, even the one which would usually incline PacMan to eat pills. So he just stands there and eventually tries to run from ghosts but never tries to finish a level.
I speculate that this happens when he loses in training, that the negative reward is propagated through the system and since the potential number of ghosts can be greater than one, this has a heavy bearing on the weights, causing everything to become very negative and the system can't "recover" from this.
I confirmed this by adjusting the feature extractor to only scale the #-of-ghosts-one-step-away feature and then PacMan manages to get a much better result
In retrospect this question is now more mathsy and might fit better on another stackexchange.