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.

Filter by
Sorted by
Tagged with
2
votes
0answers
14 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 ?
0
votes
0answers
17 views

Which loss function should I use to train DDGP with multiple q values, one for each of the output dimensions?

I'm trying to come up with a loss function for the case, in DDPG, where we have as many outputs from the critic as there are from the actor. So, there will be one Q value for each dimension in the ...
2
votes
1answer
18 views

How does DDPG algorithm know about my action mapping function?

I am using DDPG to solve a RL problem. The action space is given by the Cartesian product $[0,20]^4\times[0,6]^4$. The actor is implemented as a deep neural network ...
4
votes
1answer
58 views

How does AlphaZero's MCTS work when starting from the root node?

From the AlphaGo Zero paper, during MCTS, statistics for each new node are initialized as such: ${N(s_L, a) = 0, W (s_L, a) = 0, Q(s_L, a) = 0, P (s_L, a) = p_a}$. The PUCT algorithm for selecting ...
2
votes
0answers
24 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
35 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
1answer
31 views

Is there a training data capacity limit for AlphaZero (Chess)?

In AlphaZero, we collect ($s_t, \pi_t, z_t$) tuples from self-play, where $s_t$ is the board state, $\pi_t$ is the policy, and $z_t$ is the reward from winning/losing the game. In other DeepRL off-...
3
votes
0answers
64 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, ...
1
vote
1answer
45 views

How to build a Neural Network to approximate the Q-function?

I am learning reinforcement learning with Q-learning using online resources, like blog posts, youtube videos, and books. At this point, I have learned the underpinning concepts of reinforcement ...
0
votes
0answers
32 views

In AlphaZero, which features are one-hot encoded and which are single real-valued?

From the AlphaZero paper, the caption of Table S1 (p. 13) Table S1: Input features used by AlphaZero in Go, Chess and Shogi respectively. The first set of features are repeated for each position in a ...
1
vote
0answers
38 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 ...
2
votes
1answer
41 views

In AlphaZero, do we need to store the data of terminal states?

I have a question about the training data used during the update/back-propagation step of the neural network in AlphaZero. From the paper: The data for each time-step $t$ is stored as ($s_t, \pi_t, ...
1
vote
0answers
23 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 ...
2
votes
0answers
21 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 ...
3
votes
1answer
61 views

How to build a DQN agent with state and action being arrays?

I have a Reinforcement-Learning environment where the state is an array of 0s and 1s with length equals to the number of users the agent must satisfy (11 users). The agent must choose one of 12 ...
3
votes
0answers
29 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: ...
1
vote
0answers
25 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
1answer
39 views

Why are shallow networks so prevalent in RL?

In deep learning, using more layers in a neural network adds the capacity to capture more features. In most RL papers, their experiments use a 2 layer neural network. Learning to Reset, Constrained ...
1
vote
0answers
70 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 ...
2
votes
0answers
25 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 ...
0
votes
0answers
27 views

What are the implications of storing the alternative situation (that could have been experienced) in the replay buffer?

Consider an environment where there are 2 outcomes (e.g. dead and alive) and a discrete set of actions. For example, a game where the agent has 2 guns $A$ and $B$ to shoot a monster (the monster dies ...
1
vote
1answer
40 views

Initialising DQN with weights from imitation learning rather than policy gradient network

In AlphaGo, the authors initialised a policy gradient network with weights trained from imitation learning. I believe this gives it a very good starting policy for the policy gradient network. the ...
3
votes
1answer
49 views

How is MuZero's second binary plane for chess defined?

From the MuZero paper (Appendix E, page 13): In chess, 8 planes are used to encode the action. The first one-hot plane encodes which position the piece was moved from. The next two planes encode ...
3
votes
1answer
45 views

DQN layers when state space and action space are multi dimensional

I have built my own RL environment, where a state is composed of two elements: the agent's position and a matrix of 0s and 1s (1 if a user has requested a service from the agent, 0 otherwise); an ...
1
vote
1answer
44 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 ...
0
votes
0answers
18 views

Need suggestion for Reinforcement Learning based visual landing system for quadcopters (UAVs)

I have deep interest in quadcopters. I need ideas for designing of an experiment. I have a programmable quadcopter. I can autonomously land it on a staitonary landing pad with a vision algorithm. ...
1
vote
2answers
116 views

Why not use the target network in DQN as the predictor after training

Target network in DQN is known to make the network more stable, and the loss is like "how good I'm now compared to using the target". What I don't understand is, if the target network is the ...
1
vote
0answers
47 views

How does one stack multiple observations in the input layer of a convolutional neural network?

The paper, Deep Recurrent Q-Learning for Partially Observable MDPs, talks about stacking multiple observations in the input of a convolutional neural network. How does this exactly work? Do the ...
2
votes
1answer
63 views

Why is tree search/planning used in reinforcement learning?

In AlphaGo Zero, MCTS is used along with policy networks. Some sources say MCTS (or planning in general) increases the sample efficiency. Assumed the transition model is known and the computational ...
2
votes
2answers
95 views

How should I define the reward function to solve the Wumpus game with deep Q-learning?

I'm writing a DQN agent for the Wumpus game. Is the reward function to train the Q-networks (target network and policy) the same as the score of the game, i.e. +1000 for picking up gold, -1000 for ...
1
vote
0answers
41 views

How to design an observation(state) space for a simple `Rock-Paper-Scissor` game?

For weeks I've been working with this toy game of Rock-Paper-Scissor. I want to use a PPO agent learn to beat a computer ...
0
votes
1answer
26 views

Why would the reward of A3C with LSTM suddenly drop off after many episodes?

I am training an A3C with stacked LSTM. During initial training, my model was giving descent +ve reward. However, after many episodes, its reward just goes to zero and is continuing for a long time. ...
0
votes
0answers
21 views

Reward and loss follow the same shape in DQN

If the accumulated reward increases, the loss increases and vice versa. This is a strange behaviour. See the figure below for an example. What is the possibility of having this behaviour in DQN? I ...
0
votes
0answers
38 views

Does Multi-Agent Deep Deterministic Policy Gradient also work with discrete action spaces?

I would like to ask if Multi-Agent Deep Deterministic Policy Gradient (MADDPG) works fine with discrete action space. DDPG works only with continuous action space, but I have read that MADDPG can also ...
1
vote
0answers
40 views

Why would DDPG with Hindsight Experience Replay not converge?

I am trying to train a DDPG agent augmented with Hindsight Experience Replay (HER) to solve the KukaGymEnv environment. The actor and critic are simple neural networks with two hidden layers (as in ...
1
vote
0answers
23 views

Is it feasible to train a DQN with thousands of input ports?

I designed a DQN architecture for some problem. The problem has a parameter $m$ as the number of clients. In my situation, $m$ is large, $m\in\{100,200,\ldots,1000\}$. For this situation, the number ...
3
votes
1answer
71 views

How can I fix jerky movement in a continuous action space

I am training an agent to do object avoidance. The agent has control over its steering angle and its speed. The steering angle and speed are normalized in a $[−1,1]$ range, where the sign encodes ...
3
votes
1answer
52 views

How we are calculating average reward ($r(\pi)$) if the policy changes over time?

In the average reward setting the quality of a policy is defined as: $$ r(\pi) = \lim_{h\to\infty}\frac{1}{h} \sum_{j=1}^{h}E[R_j] $$ When we reach the steady state distribution we can write the above ...
2
votes
1answer
43 views

What is the optimal exploration-exploitation trade-off in Q*bert?

I am training an RL agent with Deep Q-learning + Experience Replay on the Q*bert Atari environment. After 400,000 frames, my agent appears to have learned strategic information about the game, but ...
4
votes
1answer
63 views

Is there a logical method of deducing an optimal batch size when training a Deep Q-learning agent with experience replay?

I am training an RL agent using Deep-Q learning with experience replay. At each frame, I am currently sampling 32 random transitions from a queue which stores a maximum of 20000 and training as ...
1
vote
0answers
47 views

Why scaling reward drastically affects performance?

I have devised an gridworld-like environment where a RL agent is tasked to cover all the blank squares by passing through them. Possible actions are up, down, left, right. The reward scheme is the ...
1
vote
1answer
61 views

What are the variables that need to be saved and loaded, so that a DQN model starts where it left off?

TensorFlow allows users to save the weights and the model architecture, however, that will be insufficient unless the values of certain other variables are also stored. For instance, in DQN, if $\...
1
vote
1answer
69 views

How is exponential moving average computed in deep Q networks?

In normal Q-learning, the update rule is an implementation of the exponential moving average, which then converges to the optimal true Q values. However, looking at DQN, how exactly is the exponential ...
4
votes
1answer
314 views

How does the Ornstein-Uhlenbeck process work, and how it is used in DDPG?

In section 3 of the paper Continuous control with deep reinforcement learning, the authors write As detailed in the supplementary materials we used an Ornstein-Uhlenbeck process (Uhlenbeck & ...
3
votes
1answer
87 views

How to compute the target for double Q-learning update step?

I've already read the original paper about double DQN but I do not find a clear and practical explanation of how the target $y$ is computed, so here's how I interpreted the method (let's say I have 3 ...
1
vote
1answer
42 views

Why are Target Networks used in Deep Q-Learning as opposed to the Expected Value equation?

I understand we use a target network because it helps resolve issues regarding stability, however, that's not what I'm here to ask. What I would like to understand is why a target network is used as a ...
0
votes
0answers
30 views

Select training data for episodic reinforcement learning (stock trading agent)

I playing around with a stock trading agent trained via (deep) reinforcement learning, including memory replay. The agent is trained for 1000 episodes, where each episodes consists of 180 timesteps (e....
0
votes
2answers
129 views

What are some (deep) reinforcement learning books for beginners? [duplicate]

What are some books on reinforcement learning (RL) and deep RL for beginners? I'm looking for something as friendly as the head first series, that breaks down every single thing.
1
vote
0answers
42 views

How does DQN convergence work in reinforcement learning

In supervised learning we have an unbiased target value, but in reinforcement learning this isn’t the case The network predicts its own target value, now how exactly does it converge if the network ...
1
vote
0answers
34 views

How to combine two differently equally important signals into the reward function, that have different scales?

I have two signals that I want to use to model my reward. The first one is the CPU TIME: running mean from this diagram: The second one is the MAX RESIDUAL from this diagram: Since they are both ...

1
2 3 4 5