Questions tagged [deep-rl]

For questions related to deep reinforcement learning (DRL), that is, RL combined with deep learning. More precisely, deep neural networks are used to represent e.g. value functions or policies.

102 questions with no upvoted or accepted answers
Filter by
Sorted by
Tagged with
5
votes
1answer
2k views

How to deal with different actions for different states of the environment?

I'm new to this AI/Machine Learning and was playing around with OpenAI Gym a bit. When looking through the environments, I came across the Blackjack-v0 environment, ...
4
votes
1answer
108 views

What could be the cause of the drop in the reward in A3C?

The mean episodic reward is generally increasing, but it has spontaneous drops, and I'm not sure of their cause. The problem has a sparse reward, batch size=2000, <...
4
votes
0answers
382 views

What could be causing the drastic performance drop of the DQN model on the Pong environment?

I am running a basic DQN (Deep Q-Network) on the Pong environment. Not a CNN, just a 3 layer linear neural net with ReLUs. It seems to work for the most part, but at some point, my model suffers from ...
4
votes
0answers
124 views

How did the OpenAI 5 for Dota concatenate units?

I am no expert in the field of AI so I apologize if this is a simple/easy question. I was trying to implement a network similar to OpenAI's for another game and I noticed that I did not fully ...
3
votes
0answers
64 views

Where does this variation of the importance sampling weight come from?

I have seeing a variation in importance sampling (IS) in Prioritized Experience Replay (PER) in some implementations regarding the original paper approach stated as (in section 3.4): $$ w_{i}=\left(\...
3
votes
0answers
37 views

DQN Agent with a 2D matrix as input in Keras

I have a Reinforcement Learning environment where the state is a 2D matrix with 0s and 1s (only one column with the value of 1 in each row). Example: ...
3
votes
0answers
82 views

How to take actions at each episode and within each step of the episode in deep Q learning?

In deep Q learning, we execute the algorithm for each episode, and for each step within an episode, we take an action and record a reward. I have a situation where my action is 2-tuple $a=(a_1,a_2)$. ...
3
votes
0answers
59 views

Understanding the TensorFlow implementation of the policy gradient method

I was trying to understand the implementation of a basic policy gradient (REINFORCE) method using TensorFlow. I think I got almost everything. The only thing that still bothers me is the loss function ...
3
votes
0answers
34 views

How does normalization of the inputs work in the context of PPO?

What does the normalization of the inputs mean in the context of PPO? At each time step of an episode, I only know the values of this time step and of the previous ones, if I take track of them. This ...
3
votes
0answers
129 views

How to correctly implement self-play with DQN?

I have an environment where an agent faces an equal opponent, and while I've achieved OK performance implementing DQN and treating the opponent as a part of the environment, I think performance would ...
3
votes
0answers
52 views

Is this a good approach to solving Atari's “Montezuma's Revenge”?

I'm new to Reinforcement Learning. For an internship, I am currently training Atari's "Montezuma's Revenge" using a double Deep Q-Network with Hindsight Experience Replay (HER) (see also ...
3
votes
1answer
76 views

How does the optimization process in hindsight experience replay exactly work?

I was reading the following research paper Hindsight Experience Replay. This is the paper that introduces a concept called Hindsight Experience Replay (HER), which basically attempts to alleviate the ...
3
votes
0answers
43 views

How to deal with nonstationary rewards in asymmetric self-play reinforcement learning?

Suppose we're training two agents to play an asymmetric game from scratch using self play (like Zerg vs. Protoss in Starcraft). During training one of the agents can become stronger (discover a good ...
3
votes
0answers
169 views

Deep Q-Network (DQN) to learn the game 2048

I am trying to build a Deep Q-Network (DQN) agent that can learn to play the game 2048. I am orientating myself on other programs and articles that are based on the game snake and it worked well (...
3
votes
0answers
29 views

Training a reinforcement learning model with multiple images

I am tentatively trying to train a deep reinforcement learning model the maze escaping task, and each time it takes one image as the input (e.g., a different "maze"). Suppose I have about $10K$ ...
3
votes
0answers
193 views

DQN Agent not learning anymore - what can I do to fix this?

I am trying to use Deep-Q-Learning to learn an ANN which controls a 7-DOF robotic arm. The robotic arm must avoid an obstacle and reach a target. I have implemented a number of state-of-art ...
3
votes
0answers
49 views

Deep Q-Learning agent poor performing actions. Need help optimizing

I'm trying to make deep q-learning agent from https://keon.io/deep-q-learning My environment looks like this: https://imgur.com/a/OnbiCtV As you can see my agent is a circle and there is one gray ...
3
votes
0answers
170 views

What can be considered a deep recurrent neural network?

In the paper Deep Recurrent Q-Learning for Partially Observable MDPs, the DRQN is described as DQN with the first post-convolutional fully-connected layer replaced by a recurrent LSTM. I have DQN ...
3
votes
1answer
112 views

Why Q2 is a more or less independant estimate in Twin Delayed DDPG (TD3)?

Twin Delayed Deep Deterministic (TD3) policy gradient is inspired by both double Q-learning and double DQN. In double Q-learning, I understand that Q1 and Q2 are independent because they are trained ...
2
votes
0answers
17 views

Can DQN outperform DoubleDQN?

I found a similar post about this issue, but unfortunately I did not find a proper answer. Are there any references where DQN is better than DoubleDQN, that is DoubleDQN does not improve DQN ?
2
votes
0answers
67 views

Policy gradient: Does it use the Markov property?

To derive the policy gradient, we start by writing the equation for the probability of a certain trajectory (e.g. see spinningup tutorial): $$ \begin{align} P_\theta(\tau) &= P_\theta(s_0, a_0, ...
2
votes
0answers
23 views

CNN for a DQN agent with a 2D matrix state and action as a 2D matrix

I have a custom environment, where the state is a 2D matrix of 11 rows (equals to number of users to satisfy) and 3 columns. Each column can take the value of either 0 or 1, and in each row, there can ...
2
votes
0answers
27 views

Is Deep SARSA learning a feasible approach?

I noticed that SARSA has been rarely used in the deep RL setting. Usually, the training for DQN is done off-policy. I think one of the major reasons for this is due to greater sample efficiency in ...
2
votes
0answers
64 views

If the minimum Q value is decreasing and the maximum Q value increasing, is this a sign that dueling double DQN is diverging?

I'm training a dueling double DQN agent with prioritized replay buffer and notice that the min Q values are decreasing, while the max Q values are increasing. Is this a sign that it is diverging? ...
2
votes
0answers
76 views

How can the target rely on untrained parameters?

I'm trying to understand DQN. I understand where the loss function comes from. I'm just unsure about why the target function works in practice. Given the loss function $$ L_i(\theta_i) = [(y_i - Q(s,a;...
2
votes
0answers
44 views

Are the final states not being updated in this $n$-step Q-Learning algorithm?

I am reading this paper and in algorithm 3 they describe an $n$-step Q-Learning algorithm. Below is the pseudo-code. From this pseudo-code, it looks as though the final tuples that they would ...
2
votes
0answers
62 views

How to prevent deep Q-learning algorithms to overfit?

I have recently solved the Cartpole problem using double deep Q-learning. When I saw how the agent was doing, it used to go right every time, never left, and it did similar actions all the time. Did ...
2
votes
0answers
35 views

Can you find another reason for sample inefficiency of model-free on-policy Deep Reinforcement Learning?

The following mindmap gives an overview of multiple reasons for sample inefficiency. The list is definitely not complete. Can you see another reason not mentioned so far? Some related links: ...
2
votes
0answers
41 views

Does the concept of validation loss apply to training deep Q networks?

In deep learning, the concept of validation loss is to ensure that the model being trained is not currently overfitting the data. Is there a similar concept of overfitting in deep q learning? Given ...
2
votes
0answers
34 views

How was the DQN trained to play many games?

Some people claim that DQN was used to play many Atari games. But what actually happened? Was DQN trained only once (with some data from all games) or was it trained separately for each game? What was ...
2
votes
0answers
417 views

How much time does it take to train DQN on Atari environment?

I am trying to build a DQN model for the Atari Pong game, but I am not sure whether the model is learning at all. I am using the architecture described in the paper Playing Atari with Deep ...
2
votes
0answers
80 views

How does adding noise to the action in DDPG help in learning?

I can't understand how playing with the action generated by the actor network in DDPG by adding the noise term helps in exploration.
2
votes
0answers
52 views

Torch CNN not training

I am completely new to CNN's, and I do not quite know how to design or use them efficiently. That being said, I am attempting to build a CNN that learns to play Pac-man with reinforcement learning. I ...
2
votes
0answers
51 views

How does the TRPO surrogate loss account for the error in the policy?

In the Trust Region Policy Optimization (TRPO) paper, on page 10, it is stated An informal overview is as follows. Our proof relies on the notion of coupling, where we jointly define the ...
2
votes
0answers
227 views

Why overfitting is bad in DQN?

It is mentioned by Fu 2019 that overfitting might have a negative effect on training DQN. They showed that with either early stopping or experience replay this effect could be reduced. The first is ...
2
votes
0answers
130 views

Why experience reply memory in DQN instead of a RNN memory?

I was trying to implement a DQN without experience reply memory, and the agent is not learning anything at all. I know from readings that experience reply is used for stabilizing gradients. But how ...
2
votes
0answers
43 views

How does Hindsight Experience Replay cope with multiple goals?

What if there are multiple goals? For example, let's consider the bit-flipping environment as described in the paper HER with one small change: now, the goal is not some specific configuration, but ...
1
vote
0answers
40 views

How should I simulate this Markov Decision Process?

I am working on solving a problem on nodes in a graph communicating with each other. They try to estimate a central state using Kalman consensus filter, with the connections described by the graph's ...
1
vote
0answers
28 views

Clarifying representation of Neural Nerwork input for Chess Alpha Zero

In the Alpha Zero paper (https://arxiv.org/pdf/1712.01815.pdf) page 13, the input for the NN is described. In the beggining of the page, the authors state that: "The input to the Neural Network ...
1
vote
0answers
19 views

Is (log-)standard deviation learned in TRPO and PPO or fixed instead?

After having read Williams (1992), where it was suggested that actually both the mean and standard deviation can be learned while training a REINFORCE algorithm on generating continuous output values, ...
1
vote
0answers
30 views

Confusion about computing policy gradient with automatic differentiation ( material from Berkeley CS285)

I am taking Berkeley’s CS285 via self-study. On this particular lecture regarding Policy Gradient, I am very confused about the inconsistency between the concept explanation and the demonstration of ...
1
vote
0answers
27 views

If one of the inputs to a neural network (that represents a policy) is noisy and degrades the performance, would this architecture solve the issue?

I'm using genetic algorithms to train deep reinforcement learning (DRL) agents, similarly to what was done in this paper. DRL policies are therefore represented by deep neural networks, which map ...
1
vote
0answers
20 views

Gradual decrease in performance of a DDPG agent

I'm trying to solve the OpenAI's CarRacing-v0 environment with the DDPG algorithm. I've observed that after a period of learning, the agent's performance starts to deteriorate slowly. For some ...
1
vote
0answers
29 views

Are inputs into AlphaZero the same during the evaluate step in MCTS and during test time?

From the AlphaZero paper: The input to the neural network is an N × N × (M T + L) image stack that represents state using a concatenation of T sets of M planes of size N × N . Each set of planes ...
1
vote
0answers
37 views

DDQN Agent in Othello (Reversi) game struggle to learn

This is my first question on this forum and I would like to welcome everyone. I am trying to implement DDQN Agent playing Othello (Reversi) game. I have tried multiple things but the agent which seems ...
1
vote
0answers
47 views

Stack of Planes as the Action Space Representation for AlphaZero (Chess)

I have a question regarding the action space of the policy network used in AlphaZero. From the paper: We represent the policy π(a|s) by a 8 × 8 × 73 stack of planes encoding a probability ...
1
vote
0answers
47 views

What are the differences between Proximal Policy Optimization versions PPO1 and PPO2?

When Proximal Policy Optimization (PPO) was released, it was accompanied by a paper describing it. Later, the authors at OpenAI introduced a second version of PPO, called PPO2 (whereas the original ...
1
vote
0answers
26 views

How to compute the Retrace target for multi-step off-policy Reinforcement Learning?

I am implementing the A3C algorithm and I want to add off-policy training using Retrace but I am having some trouble understanding how to compute the retrace target. Retrace is used in combination ...
1
vote
0answers
71 views

Does the order in which the features are concatenated to create the state (or observation) matter?

I'm experimenting with an RL agent that interacts with the following environment. The learning algorithm is double DQN. The neural network represents the function from state to action. It's build with ...
1
vote
1answer
49 views

Why do we minimise the loss between the target Q values and 'local' Q values?

I have a question regarding the loss function of target networks and current (online) networks. I understand the action value function. What I am unsure about is why we seek to minimise the loss ...