Questions tagged [reinforcement-learning]

For questions related to reinforcement learning, i.e. a machine learning technique where we imagine an agent that interacts with an environment (composed of states) in time steps by taking actions and receiving rewards (or reinforcements), then, based on these interactions, the agent tries to find a policy (i.e. a behavioural strategy) that maximizes the cumulative reward (in the long run), so the goal of the agent is to maximize the reward.

Filter by
Sorted by
Tagged with
2
votes
1answer
39 views

How does sharing parameters between the policy and value functions help in PPO?

The PPO objective may include a value function error term when parameters are shared between the policy and value functions. How does this help, and when to use a neural network architecture that ...
3
votes
1answer
98 views

What is the probability of selecting the greedy action in a 0.5-greedy selection method for the 2-armed bandit problem?

I'm new to reinforcement learning and I'm going through Sutton and Barto. Exercise 2.1 states the following: In $\varepsilon$-greedy action selection, for the case of two actions and $\varepsilon=0.5$...
4
votes
1answer
127 views

What is the effect of parallel environments in reinforcement learning?

Do parallel environments improve the agent's ability to learn or does it not really make a difference? Specifically, I am using PPO, but I think this applies across the board to other algorithms too.
4
votes
3answers
108 views

What exactly are partially observable environments?

I have trouble understanding the meaning of partially observable environments. Here's my doubt. According to what I understand, the state of the environment is what precisely determines the next state ...
0
votes
1answer
29 views

Converging to a wrong optimal policy if the agent is given more choices

I am a bit new to Reinforcement learning. So, I am extremely sorry if I am asking something obvious. I have written a small piece of code to find the optimal policy for a 5x5 grid problem. Scenario 1....
0
votes
0answers
29 views

What is the relationship between buffer size and actor loss in DDPG?

Actor loss in the DDPG algorithm is: critic_value = critic_model([state_batch, action_batch]) actor_loss = -tf.math.reduce_mean(critic_value) As I was trying to ...
1
vote
1answer
58 views

What happens with policy gradient methods if rewards are differentiable?

I would like some help with understanding why there is no explicit flow of information from the reward gradient to the parameters of the policy in policy gradient methods. What I mean is the following,...
1
vote
1answer
40 views

Why does the Atari Gym Amidar environment only move after a certain number of episodes? [closed]

When I try to run Amidar even without RL code, I cannot get the environment to move immediately. It takes about 100 steps before the game actually starts moving. I use the following simple code to ...
4
votes
1answer
185 views

Are policy gradient methods good for large discrete action spaces?

I have seen this question asked primarily in the context of continuous action spaces. I have a large action space (~2-4k discrete actions) for my custom environment that I cannot reduce down further: ...
2
votes
1answer
36 views

Is there a multi-agent deep reinforcement learning algorithm which is for environments with only discrete action spaces (Not hybrid)?

Is there a multi-agent deep reinforcement learning algorithm which is for environments with only discrete action spaces (Not hybrid) and have centralized training? I have been looking for algorithms, (...
0
votes
0answers
25 views

Task/Process Scheduling Using DQN: Agent Does Not Learn

I have a problem statement where a couple of smartphones, for example inside a shopping mall, can migrate their time-consuming tasks/processes like image processing to an edge server located nearby. ...
2
votes
1answer
54 views

In off-policy MC control algorithm by Sutton & Barto, why do we perform a last update when sample action is inconsistent with target policy?

I have a question about the $W$ term in the off-policy MC control algorithm on Page 111 of Sutton & Barto. I have also included it in the figure below. My question: shouldn't the check $A_{t} = \...
1
vote
1answer
33 views

Understanding Generalized Advantage Estimate in reinforcement learning

I was reading the paper on Generalized Advantage Estimate. It first introduces a generalized form of policy gradient equation without involving $\gamma$ and then it says the following: We will ...
1
vote
1answer
55 views

How should I interpret the surrogate and mean_noise_std plots of training a PPO model (from the Nvidia's Isaac gym)?

I am currently using the PPO method from the Nvidia's Isaac gym to train an agent for my robot. Below, you can see the plot which corresponds to a training process. I know that something is massively ...
1
vote
1answer
49 views

When showing that the policy improvement theorem applies to MC control, why is $q_{\pi_{k}}\left(s, \pi_{k}(s)\right) \geq v_{\pi_{k}}(s)$ true?

When discussing why the policy improvement theorem is true, when we do Monte Carlo control by updating greedily, it says on page 98 of Sutton and Barto's book (2nd edition) that: $$ \begin{aligned} ...
2
votes
0answers
96 views

Update Rule with Deep Q-Learning (DQN) for 2-player games

I am wondering how to correctly implement the DQN algorithm for two-player games such as Tic Tac Toe and Connect 4. While my algorithm is mastering Tic Tac Toe relatively quickly, I cannot get great ...
-2
votes
1answer
55 views

Policy Gradient ( Advantage actor-critic) for multiple simultaneous continuous actions

i'm trying to solve a problem in which i need to carry out reinforcement learning with multiple simultaneous actions in continuous action space . i checked the multiagent structure; however, im trying ...
6
votes
1answer
81 views

Understanding the On-policy state distribution for episodic tasks with $\gamma \in (0,1)$

In Sutton and Barto's Reinforcement Learning: An Introduction, section 9.2 (page 199) (here is a screenshot) describes the on-policy distribution in episodic tasks, with $\gamma =1$, as being \begin{...
2
votes
1answer
53 views

Is existence and uniqueness of state-value function at $\gamma < 1$ theoretical?

Consider the following statement from 4.1 Policy Evaluation of the first edition of Sutton and Barto's book. The existence and uniqueness of $V^{\pi}$ are guaranteed as long as either $\gamma < 1$...
1
vote
0answers
20 views

Is there any thumb rule on the cardinality of state space in order to use the parameterized function to estimate value functions?

Value functions for a given MDP can be learned in at least two ways by experience. The first way (tabular calculation) is generally used in the case of state spaces that are small enough. The second ...
3
votes
1answer
159 views

How is the state-value function expressed as a product of sums?

The state-value function for a given policy $\pi$ is given by $$\begin{align} V^{\pi}(s) &=E_{\pi}\left\{r_{t+1}+\gamma r_{t+2}+\gamma^{2} r_{t+3}+\cdots \mid s_{t}=s\right\} \\ &=E_{\pi}\...
4
votes
1answer
94 views

What are the major differences between multi-armed bandits and the other well-known algorithms (DQN, A3C, PPO, etc)?

I have studied in the past different algorithms, i.e. DQN, DDQN, REINFORCE, A3C, PPO, TRPO, so on. I am doing an internship this summer where I have to use a multi-armed bandit (MAB). I am a bit ...
0
votes
0answers
23 views

Reward firstly increase, but after more episodes, start decrease, and weights diverges

I'm making a simple deep Q learning algorithm, with cartpole-v1 env. Like you can see in chart, after many episodes the reward decrease, some possible reasons? The exploration vs axplotation algorithm ...
0
votes
0answers
12 views

DQN + HER, TD Error spiked then success rate plummets. What went wrong?

TL;DR: I trained a DQN + HER model using stable-baselines library for a custom environment. I noticed that in most runs, sometimes the TD-Error will spike and then the success rate of my model ...
0
votes
0answers
18 views

Comparison between TD(0) and MC ( or GAE )?

I'm getting started with DRL and have trouble distinguishing TD(0), MC, and GAE; and which scenarios one's better than others. Here is what I understand so far: TD(0): increment learning, can learn ...
0
votes
1answer
36 views

What would happen to an agent trained using Markov Decision Process if the goal node changes?

I was reading up a paper that did routing based on an MDP, and I was wondering because, in routing, there is a sender node and a receiver node, so if the receiver node changes (sending a message to ...
3
votes
0answers
97 views

How to deal with a moving target in the Lunar Lander environment with DDPG?

I have noticed that DDPG does rather well at solving environments with a static target. For example, the default of Lunar Lander, the flags do not change position. So the DDPG model learns how to get ...
-1
votes
1answer
35 views

Exploration for softmax should be binary or continuous softmax?

Maybe it's silly to ask but for random exploration in an RL for choosing discrete action, that in the neural network last layer softmax will be used, what random samples should we provide? binary like ...
1
vote
1answer
35 views

PPO agent for vehicle control does not learn to stop at traffic lights

I have built a custom RL environment with gym, which simulates the RL vehicle and potential vehicles in front of the RL vehicle as well as traffic lights (including ...
2
votes
0answers
33 views

suitable reference for reinforcement learning for beginners [duplicate]

I am looking for a reference book on RL for first time learners, one that is a gentle introduction and not as wordy as Sutton/Barto. I am interested in something similar to Georgia Institute of ...
0
votes
0answers
47 views

Reinforcement learning for rearranging the mobile home screen icon layout: what inputs/states do I need to pass into the algorithm?

I have a problem where I need to rearrange a particular user's mobile home screen icon layout. Let's say that the social media app usage of a user is high compared to other app usage. So I need the ...
2
votes
0answers
41 views

Why isn't RL considered a continual learning strategy itself?

I have read about methods that apply continual learning strategies to reinforcement learning. Since reinforcement learning also learns step by step (i.e., task by task, in a sense) during the training ...
2
votes
0answers
43 views

Is there any reasonable notion of regret for infinite horizon discounted MDPs?

I am thinking about episodic MDPs. Usually, in episodic MDPs, it seems that we have a finite fixed horizon per episode and no discount factor. Then, a very intuitive notion of regret after $T$ ...
3
votes
1answer
80 views

In DQN, would it be cheaper to have $N$ neural networks with a single real-valued output, one for each of the $N$ actions?

In the classical examples of deep q-learning, I often see neural networks in which the input represents the state of the agent, while the output is a tuple with all the values of $Q(s, a)$ predicted ...
0
votes
0answers
36 views

Should the actor and critic share a common feature extraction neural network?

In an environment with image observations, if we use an actor-critic method to find a good policy, commonly, we will use a feature extraction neural network, such as ResNet, to extract the information ...
1
vote
1answer
52 views

How to scale all positive continuous reward?

My RL project has all positive continuous rewards for every step and the goal is to have the maximum cumulative reward (episodic reward). The problem is that the rewards are too close and all between ...
1
vote
0answers
88 views

PPO in continuous control not working

I have PPO agent for discrete action space for LunarLander-v2 env in gym and it works well. However, when i am trying to solve continuous version of the same env - <...
0
votes
0answers
21 views

What are the state-of-the-art learning algorithms for contextual bandits with stochastic rewards

I am building a solution for an environment with stochastic rewards in an online setting. I am wondering what the state of the art is in this setting. Is it $\epsilon$-greedy (with logistic regression)...
1
vote
1answer
25 views

Why is it that the state visitation frequency equals the sum of state visitation frequency from initial time step to the horizon?

In the maximum entropy inverse reinforcement learning paper, Ziebart et al. show that the state visitation frequency $\rho(s)$ of a state $s$ can be computed as $$ \rho_{\pi}(s) = \sum_{t}^{T} P(s_t=s|...
2
votes
1answer
66 views

Difference in UCB performance when scaling the rewards

I notice the following behavior when running experiments with $\epsilon$-greedy and UCB1. If the reward is kept binary (0 or 1) both algorithm's performances are on par with each other. However, if I ...
0
votes
0answers
7 views

Update state representation based on agent action

I am working on a project where I have to train a RL agent which will simulate Loan repayment track of a customer's loan based on his features derived from his credit profile (state vector). I am ...
2
votes
2answers
112 views

Why does Alpha Zero's Neural Network flip the board to be oriented towards the current player?

While reading the AlphaZero paper in preparation to code my own RL algorithm to play Chess decently well, I saw that the "The board is oriented to the perspective of the current player." I ...
1
vote
1answer
48 views

What would be the importance sampling ratio for off-policy TD learning control using Q values?

The off-policy TD learning control using state value function from page 34 of David Silver's RL lecture is: $$ V(S_t) \leftarrow V(S_t) + \alpha \left( \frac{ \pi(A_t|S_t)}{\mu (A_t|S_t)} (R_{t+1} + \...
-1
votes
1answer
25 views

Using states (features) and actions from a heuristic model to estimate the value function of a reinforcement learning agent [closed]

new to RL here. As far as i understood from RL courses, that there is two sides of reinforcement learning. Policy Evaluation, which is the task of knowing the value function for certain policy. and ...
1
vote
0answers
31 views

Is this a supervised or reinforcement learning problem, and which algorithm should I use to solve it?

I have a time series data with a little unusual cost/reward function (I haven't seen it before) The model must predict a $Y$ value for any $X(t)$. The reward is computed as follows. The model will ...
3
votes
1answer
138 views

How does the Alpha Zero's move encoding work?

I am a beginner in AI. I'm trying to train a multi-agent RL algorithm to play chess. One issue that I ran into was representing the action space (legal moves/or honestly just moves in general) ...
0
votes
0answers
51 views

Open AI Taxi - Agent fails to learn an effective policy

I'm trying to solve the openai gym taxi problem (v3) using deep q learning. I've already had some success with the q-table approach, but for the life of me cannot manage to train a NN to learn a ...
0
votes
0answers
33 views

Any RL approaches for this 2D space optimisation problem?

I have a list of rectangles, they are in certain order in 2D at the beginning. The task is to move them to get the boundary (rectangular) of the minimal area. It's OK to push off the dotted border as ...
5
votes
1answer
205 views

Does the policy iteration convergence hold for finite-horizon MDP?

Most RL books (Sutton & Barto, Bertsekas, etc.) talk about policy iteration for infinite-horizon MDPs. Does the policy iteration convergence hold for finite-horizon MDP? If yes, how can we derive ...
0
votes
0answers
20 views

Can I use binary numbers and plain floats in the same neural network input for reinforcement learning?

I'm modeling a system whose configuration can be represented by a binary array ([1,0,0,1] or [0,1,0,0], for example), and the agent can move on a 2D space (thus having 3 DOF), and the action the agent ...

1 2
3
4 5
36