Questions tagged [q-learning]

A popular reinforcement learning algorithm, an instance of TD (temporal difference) learning.

Filter by
Sorted by
Tagged with
1
vote
0answers
23 views

Bellman Optimality Operator fixed point

I'm reading Szepesvári's book on RL. My question is concerning the proof of Theorem A.10 (p. 71). Theorem Let $V$ be the fixed point of $T^∗$ and assume that there is policy $π$ which is greedy w.r.t ...
0
votes
1answer
19 views

How is the Q value translated to action in DRL framework presented in Gu, Shixiang et al (2016)

I was reading the paper Deep Reinforcement Learning for Robotic Manipulation with Asynchronous Off-Policy Updates and I kind of understand everything. The results also seem very interesting, enabling ...
2
votes
1answer
39 views

Policy improvement in SARSA and Q learning

I have a rather trivial doubt in SARSA and Q learning. Looking at the pseudocode of the two algorithms in Sutton&Barto book, I see the policy improvement step is missing. How will I get the ...
1
vote
0answers
10 views

Q Learning Target and Prediciton

The following formula is unclear to me That Q(s,a) is my prediction make sense to me. But why is (r+ Vopt(s')) my target. s' is the successor. So my prediction for Q(s,a) should be near to Q(s',a)? ...
0
votes
0answers
34 views

Q-learning convergence with stochastic reward function

Every proof for convergence of Q-learning I can find assumes that the reward is a function $r(s, a, s')$ i.e. deterministic. However, MDPs are often defined with a stochastic reward, as exemplified in ...
0
votes
0answers
93 views

Why Does Double-Q Learning Converge So Slowly?

Working through Sutton and Barto's reinforcement learning book and van Hasselt's original Double Q-Learning (2010) paper to understand Double-Q learning. In the original paper, van Hasselt uses the ...
0
votes
1answer
28 views

Deduce the Bellman equation from the Value and Q functions

I am trying to derive/deduce the bellman equation using Value and Q-functions. I came only so far with understanding it and tried it myself in Latex: Why is the $V^*$ suddenly in $Q^\pi$ function? ...
0
votes
0answers
15 views

The successor state and action with different notation in Reinforcement Learning

I am currently writing a summary of what I understand from the bellman equation, Value- and Q-functions. I am getting different ways of writing the Bellman equation: Does a' mean the same as a_{t+1}? ...
0
votes
0answers
54 views

How DynaQ behaves in stochastic world in comparison with other reinforcement learning algorithms?

I came across of implementations of a bunch of algorithms on stochastic windy gridworld. This is the graph comparing their performance: So clearly, it seems that DynaQ performs better than all other ...
0
votes
0answers
14 views

Epochs and early stopping in deep reinforcement learning

As far as I understood, in the deep reinforcement learning (DRL) literature, an epoch is a specific number of updates of the network parameters (depending on your update frequency). Is that correct? ...
0
votes
0answers
9 views

q-learning algorithm for binary classification

please how can I use a q-learning algorithm for binary classification like classify images (cat or dog) or emails to spam or not spam, please if there is any python program for classification by using ...
0
votes
1answer
38 views

q-learning for classification problem

how can I use q-learning for classification problem I see some examples about learning the agent in a game I need to use q-learning for binary classification (classify image if it is for cat or for ...
1
vote
0answers
16 views

ε-Greedy with Q learning / SARSA can have stochastic policy?

Hello I'm now studying Q learning and SARSA with ε-Greedy , Softmax startegies. And have a question about my readings. In my readings, when SARSA with ε-Greedy, SARSA causes value-function ...
1
vote
1answer
75 views

Why does the PyTorch tutorial on DQN define state as a difference?

I'm a master's student in EECS working my way towards understanding how DQN [0] works. I'm working towards solving the CartPole-v0 task in as few iterations as possible. First of all I implemented a ...
0
votes
0answers
19 views

Finite MDPs, Pole-Balancing and Q-Learning (RL by Sutton, Barto)

Currently I am reading Reinforcement Learning by S. Sutton and A. Barto. In chapter 3 the authors introduce the concept of Finite Markov Decision Processes. A feature of finite MDPs is that the state, ...
7
votes
1answer
178 views

MDP and Sate Value Finding?

I have a complex MDP (I think) as follows. anyone can describe me simply how the value for state $V(A)^*$ is find? First Update: really for this solved question I need a canonical answer, step by ...
2
votes
1answer
90 views

SARSA when policy is not epilon greedy

I would like to clarify a doubt that I have regarding SARSA. SARSA can be used for optimal control when the policy to take action $a$ is epsilon greedy. Suppose that the policy to take action $a$ is ...
3
votes
1answer
62 views

Association between Current state/action and the far future reward

How the Agent make the association between the current $Q_t(s_t,a_t)$ and a far future reward that by nature of my environment we get reward at least after 10-15 time steps from the action taken. If ...
0
votes
1answer
84 views

Understanding the Q-learning loss function?

Perhaps this can be explained a little more to me. I understand what's in literature but I'm struggling to understand why this is the preferred loss. If we have an agent that can move ↑↓→← and for ...
1
vote
1answer
53 views

Learning in the bit flipping environment

I'm looking at the Andrychowicz et al. paper and running through the gauntlet of implementing a DQN and then implementing a DQN with HER. In this paper, they mention a bit-flipping environment: a ...
0
votes
0answers
56 views

As epsilon decays, rewards gets worse during exploitation than exploration

I am currently trying to write learning agent from the "Human Level Control in DRL" Paper in Tensorflow 2.0. I've copied the recommended hyperparameters and picked the easiest environment ...
0
votes
0answers
60 views

sarsa, epxected sarsa, and qlearning convergence

After implementing all 3 algorithms for problem X, it turns out that sarsa converges significantly faster than both expected sarsa and qlearning. What causes this? Is this affected by the problem? or ...
0
votes
0answers
9 views

Q-learning with evaluation more the 1 step ahead

I am trying to make an AI play an arcade game. The game requires some complex decisions and the agent is not rewarded instantly after performing an action. His projectiles take time to travel and some ...
1
vote
0answers
92 views

Why and under what conditions does Q learning converge?

I am looking for a modern proof on why Q learning converges in the tabular setting. I've skimmed the original proof by Dayan and Watkins and I have to say that the terminology and approach are a bit ...
2
votes
1answer
35 views

Does gradient descent work for tabular Q learning?

Suppose I have a tabular Q learning problem such as grid-world. Let our loss be defined as, $$\hat{L}(Q)=0.5(Q(s,a)-(r+\gamma\max_{a'}{Q(s',a')}))^2$$ Then $Q_{k+1}(s,a) = Q_k(s,a) - \eta \nabla \hat {...
1
vote
2answers
109 views

In Reinforcement Learning (DQN), is there a way to constrain/penalise the model so that it doesn't take a different action very often?

The RL model I am building is one form of DQN. Its internal network is a regular deep NN. In the application I am looking at, there is a cost for taking a different action (compared to the previous ...
1
vote
0answers
44 views

How to define a Reinforcement Learning episode when dealing with dataset of customer's features?

so one of friends sent me some problem he was working on lately and I couldn't help but to wonder how could it be solved using Q-learning. The statement is as follows : Given the following datasets, ...
1
vote
0answers
23 views

Equivalence of state-action value and value function?

In Reinforcement Learning (RL) two of the most widespread measures of performance for a policy $\pi$ are the value function $$V^\pi\left(s\right)=\mathbb{E}_\pi\left[\sum_{i=0}^{\infty}{\gamma^ir_{t+i}...
0
votes
0answers
37 views

Dueling DQN - solving the identifiability problem

I want to implement a Dueling DQN. I know that the idea is to estimate V (scalar), the value of a given state, and A (vector) - the advantage, separately. A is the difference between V and the Q ...
2
votes
0answers
44 views

Preprocessing data for the learning step

I am currently reading "Human level control through deep reinforcement learning" and I came across the algorithm in the paper. I am confused because the algorithm uses a different notation ...
-1
votes
1answer
53 views

Deep Q learning algorithm

Ok I’m a bit confused with this code, what exactly is a time step, isn’t it like when an action is performed,it goes to the next time step, and also, the gradient descent steps is a repeat until ...
2
votes
1answer
116 views

In a DQN, can Prioritized Experience Replay actually perform worse than a regular Experience Replay?

I've written a Double DQN-based stock trading bot using mainly time series stock data. I've recently upgraded my Experience Replay(ER) code with a version of Prioritized Experience Replay (PER) ...
2
votes
2answers
121 views

Can an agent be trained in a completely random environment (the rules and actions stay the same)

I am specifically talking about the FrozenLake example from openAI gym to illustrate my question. https://gym.openai.com/envs/FrozenLake-v0/ The way I understand how Q-Learning or DQL works on the ...
0
votes
1answer
62 views

Consider more than just a single state-action pair before giving a reward in deep q learning?

TLDR If the environment is rather complex (momentum, velocity, ...), but the action space is rather easy (left, right), is it not better to take multiple decisions into account before giving a reward ...
2
votes
1answer
59 views

Q-learning update formula

I am a beginner and it is my first question. I know that Q-learning update equation is: $Q(s_t, a_t) = Q(s_t, a_t)+α(r_{t+1} +γ·max_AQ(s_{t+1}, a_t)−Q(s_t, a_t))$ But in some of the researches it is ...
0
votes
1answer
48 views

Q-Learning [Sutton]: why random variable in formula

Sutton et al. use throughout their book Reinforcement Learning capital letters to describe random variables. At page 131 they introduce Q-Learning. $Q(S_t,A_t)\leftarrow Q(S_t,A_t) + \alpha [R_{t+1} ...
1
vote
0answers
208 views

Expected SARSA, SARSA and Q-learning

I would much appreciate if you could point me in the right direction regarding this question about targets for approximate ...
1
vote
1answer
244 views

Deep Q Learning best practice

I'm new in deep q-learning and I have understood its main concepts and I'm trying to solve problems with DQL. The problem is that I don't know how to initialize some key values (AKA hyperparameters) ...
0
votes
1answer
17 views

Small difference of q-function between different actions for the same state

I am trying out reinforcement learning using Q-learning. The data come from some made-up equations so I have infinite number of data. One thing that troubles me is after I learn the Q-function, I use ...
3
votes
1answer
418 views

DQN - agent doesn't improve policy

I have a simple grid environment. The player is in the upper left corner and it's goal is to get to lower right corner. The player receives +0.2 points for moving in the direction of the goal, -0.2 ...
1
vote
1answer
663 views

Delayed Rewards in Reinforcement Learning

I have an MDP where the rewards are delayed for six steps as follows: The reward from action at time t is received when the action at time t+6 is taken. The reward from action at time t+1 is ...
0
votes
0answers
37 views

How to use data from other policies in order to find optimal policy in model-free rl?

I am struggling to understand whether experience from one policy can be used to find optimal policy. Suppose that I have gathered many data (state, action, reward, next_state) by following random ...
1
vote
1answer
701 views

Q-learning agent stucks in an infinite loop

I am simulating a mouse to find a cheese on an empty table. I randomly put a cheese on the table and let the mouse find the cheese without falling off the table. The problem is, in test part, agent ...
2
votes
1answer
49 views

Rationale behind Q-learning

I am reading Sutton Barto on Reinforcement Learning. I understand that $TD(\lambda)$ methods propose better performance than Monte Carlo methods, with TD methods combining advantages of Dynamic ...
3
votes
1answer
362 views

Reward attribution in deep q learning and texas holdem poker

I’m having issues with reward attribution in poker using deep q learning. Multiple actions will yield one reward, but the reward is only known at the end of the hand, not before. I have built a gym ...
1
vote
1answer
34 views

Why not sample action from Q values?

When collecting experience from which to estimate a Q(s,a) function, one common technique in the literature is to follow an epsilon greedy-strategy. In this strategy, the agent selects a random action ...
1
vote
1answer
449 views

How does the Dyna Q algorithm works?

I'm having a hard time trying to understand how the dyna Q algorithm works. I put the picture which helps me to understand. My questions are: What planning really means? (it's the (f) in this picture)...
0
votes
1answer
544 views

Episodic Semi-gradient Q-learning for Estimating approximation of optimal action-value function

at page 244 of Sutton and Barto book on Reinforcement Learning (book) is described the pseudocode for episodic semi-gradient Sarsa, while it is never given a pseudocode for the corresponding episodic ...
2
votes
1answer
858 views

DQN - How to feed the input of 4 still frames from a game as one single state input

I was reading this blog about Deep Q-Learning. 1- In the The input section of the blog, I wanted to know how do we feed the 4 still-frames/screenshots from the game, that represent the input state, ...
0
votes
2answers
330 views

DQN - breaking correlation between consecutive samples and random sampling

I was reading through some blogs about Deep Q-Learning (DQN), and I have 2 questions: 1- I didn't understand how breaking the correlation between consecutive samples (i.e. train the network with ...