Questions tagged [reinforcement-learning]

For questions related to learning controlled by external positive reinforcement or negative feedback signal or both, where learning and use of what has been thus far learned occur concurrently.

Filter by
Sorted by
Tagged with
1
vote
1answer
31 views

Additional (Potential) Action for Agent in MazeGrid Environment (Reinforcement Learning)

In a classic GridWorld Environment where the possible actions of an agent are (Up, Down, Left, Right), can another potential output of Action be "x amount of steps" where the agent takes 2,3,.. steps ...
1
vote
2answers
60 views

Should I use exploration strategy in Policy Gradient algorithms?

In policy gradient algorithms the output is a stochastic policy - a probability for each action. I believe that if I follow the policy (sample an action from the policy) I make use of exploration ...
1
vote
1answer
26 views

How do we reach at the formula for UCB action-selection in multi-armed bandit problem?

I came across the formula for Upper Confidence Bound Action Selection (while studying multi-armed bandit problem), which looks like: $$ A_t \dot{=} \operatorname{argmax}_a \left[ Q_t(a) + c \sqrt{ \...
0
votes
1answer
61 views

Why am I getting better performance with Thompson sampling than with UCB or $\epsilon$-greedy in a multi-armed bandit problem? [closed]

I ran a test using 3 strategies for multi-armed bandit: UCB, $\epsilon$-greedy, and Thompson sampling. The results for the rewards I got are as follows: Thompson sampling had the highest average ...
1
vote
0answers
43 views

Proof of Maximization Bias in Q-learning?

In the textbook "Reinforcement Learning: An Introduction" by Richard Sutton and Andrew Barto, the concept of Maximization Bias is introduced in section 6.7, and how Q-learning "over-estimates" action-...
0
votes
1answer
25 views

Soft Actor Critic - Losses are not converging

I'm trying to implement soft actor-critic algorithm for financial data (stock prices) and I have trouble with losses, no matter what combination of HyperParameters I enter, they are not converging, ...
0
votes
0answers
22 views

Using deep deterministic policy gradient in OpenAI Gym to solve problems with continuous actions

I am trying to do the following: Install the OpenAI baseline algorithms from the following GitHub repository: github.com/openai/baselines by following the instructions in the readme file. Train an ...
0
votes
1answer
44 views

What are the differences between SARSA and Q-learning?

From Sutton and Barto's book Reinforcement Learning (Adaptive Computation and Machine Learning series), are the following definitions: To aid my learning of RL and gain an intuition, I'm focusing on ...
2
votes
1answer
62 views

Why we don't use importance sampling in tabular Q-Learning?

Why don't we use an importance sampling ratio in Q-Learning, even though Q-Learning is an off-policy method? Importance sampling is used to calculate expectation of a random variable by using data ...
1
vote
1answer
30 views

Why is the “reward to go” replaced by Q instead of V, when transitioning from PG to actor critic methods?

While transitioning from simple policy gradient to the actor-critic algorithm, most sources begin by replacing the "reward to go" with the state-action value function (see this slide 5). I am not ...
1
vote
1answer
41 views

Solution to exercise 3.22 in the RL book by Sutton and Barto

The goal is to find an optimal deterministic policy for this MDP: There are two possible policies: left (L) and right (R). What is the optimal policy, when different discounts are used: A $\gamma =...
1
vote
1answer
35 views

What is meant by the rank of the scoring function here?

I've been reading this paper on Knowledge Graph Reasoning for Explainable Recommendation lately, and I don't understand a particular section: Specifically, the scoring function $f((r,e)|u)$ maps any ...
1
vote
1answer
135 views

How does Hindsight Experience Replay learn from unsuccessful trajectories

I am confused by how HER learns from unsuccessful trajectories. I understand that from failed trajectories it creates 'fake' goals that it can learn from. Ignoring HER for now, if in the case where ...
1
vote
2answers
32 views

Generalising performance of Q-learning agent through self-play in a two-player game (MCTS?)

I'm using Q-learning (off-policy TD-control as specified in Sutton's book on pg 131) to train an agent to play connect four. My goal is to create a strong player (superhuman performance?) purely by ...
0
votes
1answer
53 views

How do I know that the DQN has learnt an appropriate Q function?

Is there any sanity check to know whether the Q functions learnt are appropriate in deep Q networks? I know that the Q values for end states should approximate the terminal reward. However, is it ...
3
votes
2answers
130 views

What is the value of a state when there is a certain probability that agent will die after each step?

We assume infinite horizon and discount factor $\gamma = 1$. At each step, after the agent takes an action and gets its reward, there is a probability $\alpha = 0.2$, that agent will die. The assumed ...
1
vote
1answer
50 views

Why do my rewards fall using tabular Q-learning as I perform more episodes?

Using the tutorial from: SentDex - Python Programming I added Q Learning to my script that was previously just picking random actions. His script uses the MountainCar Environment so I had to amend it ...
0
votes
0answers
33 views

Model Based rl and cross entropy method with nonlinear function approximators

Pseudo code for Cross entropy method according to youtube lecture 32:55 Initialize $\mu \in R^{d}, \sigma \in R^{d}$ iteration 1,2,... Collect n samples of $\theta_{i} \sim N(\mu,diag(\sigma))$ ...
2
votes
1answer
56 views

NoisyNet DQN with default parameters not exploring

I implemented a DQN algorithm that plays OpenAIs Cartpole environment. The NN architecture consists of 3 normal linear layers that encode the state, and one noisy linear layer, that predicts the Q ...
1
vote
0answers
52 views

How is per-decision importance sampling derived in Sutton & Barto's book?

In per-decison importance sampling given in Sutton & Barto's book: Eq 5.12 $\rho_{t:T-1}R_{t+k} = \frac{\pi(A_{t}|S_{t})}{b(A_{t}|S_{t})}\frac{\pi(A_{t+1}|S_{t+1})}{b(A_{t+1}|S_{t+1})}\frac{\pi(...
1
vote
2answers
194 views

In the context of importance sampling ratio, how is the equation $\mathbb{E}\left[\rho_{t: T-1} G_{t} | S_{t}=s\right]=v_{\pi}(s)$ derived?

When reading the book by Sutton and Barto, I came across the importance sampling ratio. The first equation, I believe, describes the probability a particular sequence is obtained given the current ...
1
vote
0answers
32 views

Why does the n-step return being zero result in high variance in off policy n-step TD?

In the paragraph given between eq 7.12 and 7.13 in Sutton & Barto's book: $G_{t:h} = R_{t+1} + G_{t+1:h} , t < h < T$ where $G_{h:h} = V_{h-1}(S_h)$. (Recall that this return is used at ...
0
votes
0answers
38 views

Problem in understanding equation given for convergence of TD(n) algorithm

Given equation 7.3 of Sutton and Barto's book for convergence of TD(n): $\max_s|\mathbb{E}_\pi[G_{t:t+n}|S_t = s] - v_\pi(s)| \leqslant \gamma^n \max_s|V_{t+n-1}(s) - v_\pi(s)|$ $\textbf{PROBLEM ...
1
vote
1answer
38 views

Is it possible to prove that the target policy is better than the behavioural policy based on learned Q values?

I have retrospective data for a sort of "behaviour policy" which I will use to train a deep q network to learn a target greedy policy. After learning the Q values for this target policy, can we make ...
0
votes
2answers
84 views

When we use a neural network to approximate the Q values, is the Q target a single value?

I have two questions When we use our network to approximate our Q values, is the Q target a single value? During backpropagation, when the weights are updated, does it automatically update the Q ...
5
votes
2answers
70 views

Why state-action value function as an expected value of the return and state value function, does not need to follow policy?

I often see, the state-action value function is expressed as: $q_{\pi}(s,a)=\mathbb{E}_{\pi}[R_{t+1}+\gamma G_{t+1} | S_t=s, A_t = a] = \mathbb{E}[R_{t+1}+\gamma v_{\pi}(s') |S_t = s, A_t =a]$ Why ...
0
votes
0answers
45 views

TD-Leaf struggles at learning chess

I am currently working on implementing Giraffe chess algorithm. Following this paper, I designed a neural network similar to the one proposed by the author which I trained using TD-Leaf(lambda). The ...
0
votes
0answers
16 views

My Double DQN with Experience Replay produces a no-action decision most of the time. Why?

I've written a Double DQN-based stock trading bot using mainly time series stock data. The internal network of the Double DQN is a LSTM which handles the time series data. An Experience Replay buffer ...
3
votes
1answer
49 views

Reward Function for Racing Game

I'm busy working on a project where I'm building an agent for a racing game. In this game is a randomised map where there are speed boosts for the player to pick up and obstacles that act to slow the ...
1
vote
0answers
54 views

What is the proof that “reward-to-go” reduces variance of policy gradient?

I am following the OpenAI's spinning up tutorial Part 3: Intro to Policy Optimization. It is mentioned there that the reward-to-go reduces the variance of the policy gradient. While I understand the ...
0
votes
0answers
22 views

Can Reinforcement Learning be used for UAV waypoint control?

I want to make a drone which can follow static and dynamic waypoints. I am a total beginner in the drone field so I can't figure out that should I use Reinforcement Learning or any other learning ...
4
votes
1answer
52 views

How does policy evaluation work for continuous state space model-free approaches?

How does policy evaluation work for continuous state space model-free approaches? Theoretically, a model-based approach for the discrete state and action space can be computed via dynamic programming ...
1
vote
1answer
43 views

Do smaller loss values during DQN training produce better policies?

During the training of DQN, I noticed that the model with prioritized experience replay (PER) had a smaller loss in general compared to a DQN without PER. The mean squared loss was an order of ...
9
votes
2answers
701 views

Why is baseline conditional on state at some timestep unbiased?

In the homework for the Berkeley RL class, problem 1, it asks you to show that the policy gradient is still unbiased if the baseline subtracted is a function of the state at time step $t$. $$ \...
2
votes
1answer
86 views

Non-differentiable reward function to update a neural network

In Reinforcement Learning, when reward function is not differentiable, a policy gradient algorithm is used to update the weights of a network. In the paper Neural Architecture Search with ...
4
votes
2answers
150 views

Why are the value functions sometimes written with capital letters and other times with lower-case letters?

Why are the state-value and action-value functions are sometimes written in small letters and other times in capitals? For instance, why in the Q-learning algorithm (page 131 of Barto and Sutton's ...
2
votes
1answer
281 views

Is there any research work that attempts to combine neuroevolution with deep reinforcement learning?

Neuroevolution can be used to evolve a network's architecture (and weights, of course). Deep reinforcement learning, on the other hand, has been proven to be extremely powerful at optimising the ...
0
votes
0answers
27 views

Why in RL function approximators with recurrent structures can learn planning?

In the paper An Investigation of Model-Free Planning the authors use ConvLSTM to learn a planning function. In particular, for each input x_t at time-step ...
5
votes
1answer
1k views

How can we use linear programming to solve an MDP?

Apparently, we can solve an MDP (that is, we can find the optimal policy for a given MDP) using a linear programming formulation. What's the basic idea behind this approach? I think you should start ...
12
votes
1answer
2k views

Why does DQN require two different networks?

I was going through this implementation of DQN and I see that on line 124 and 125 two different Q networks have been initialized. From my understanding, I think one network predicts the appropriate ...
0
votes
0answers
67 views

Should the network weights converge when training Deep Q networks?

I have two sets of data, one training and one test set. I use the train set to train the deep q network model variant. I also continuously evaluate the agent Q values obtained on the test set every ...
3
votes
1answer
121 views

How do we express $q_\pi(s,a)$ as a function of $p(s',r|s,a)$ and $v_\pi(s)$?

The task (exercise 3.13 in the RL book by Sutton and Barto) is to express $q_\pi(s,a)$ as a function of $p(s',r|s,a)$ and $v_\pi(s)$. $q_\pi(s,a)$ is the action-value function, that states how good ...
3
votes
1answer
68 views

What is the input to AlphaGo's neural network?

I have been reading an article on AlphaGo and one sentence confused me a little bit, because I'm not sure what it exactly means. The article says: AlphaGo Zero only uses the black and white stones ...
1
vote
1answer
73 views

How to implement RAM versions of Atari games

I have coded the breakout RAM version, but, unfortunately, its highest reward was 5. I trained it for about 2 hours and never reached a higher score. The code is huge, so I can't paste here, but, in ...
5
votes
1answer
168 views

How do I recognise a bandit problem?

I'm having difficulty understanding the distinction between a bandit problem and a non-bandit problem. An example of the bandit problem is an agent playing $n$ slot machines with the goal of ...
2
votes
2answers
82 views

How to convert sequences of images into state in DQN?

I recently read the DQN paper titled: Playing Atari with Deep Reinforcement Learning. My basic and rough understanding of the paper is as follows: You have two neural networks; one stays frozen for a ...
14
votes
3answers
838 views

Why doesn't Q-learning converge when using function approximation?

The tabular Q-learning algorithm is guaranteed to find the optimal $Q$ function, $Q^*$, provided the following conditions (the Robbins-Monro conditions) regarding the learning rate are satisfied $\...
2
votes
0answers
29 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? ...
3
votes
1answer
49 views

Simplification of expected reward under the limit in continuous tasks

I was reading the average reward setting for continuous tasks from rich sutton's book (page 202, 2nd edition). There he perform a simplification over the expected reward under the limit approaching to ...
4
votes
1answer
48 views

How can a single sample represent the expectation in gradient temporal difference learning?

I was reading the gradient temporal difference learning version 2(GTD2) from rich Sutton's book page-246. At some point, he expressed the whole expectation using a single sample from the environment. ...

1 2
3
4 5
25