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.

714 questions with no upvoted or accepted answers
Filter by
Sorted by
Tagged with
7 votes
0 answers
210 views

Is the Bellman equation that uses sampling weighted by the Q values (instead of max) a contraction?

It is proved that the Bellman update is a contraction (1). Here is the Bellman update that is used for Q-Learning: $$Q_{t+1}(s, a) = Q_{t}(s, a) + \alpha*(r(s, a, s') + \gamma \max_{a^*} (Q_{t}(s', ...
user avatar
7 votes
0 answers
80 views

Normalizing Normal Distributions in Thompson Sampling for online Reinforcement Learning

In my implementation of Thompson Sampling (TS) for online Reinforcement Learning, my distribution for selecting $a$ is $\mathcal{N}(Q(s, a), \frac{1}{C(s,a)+1})$, where $C(s,a)$ is the number of times ...
user avatar
  • 71
7 votes
0 answers
1k views

Is there a difference in the architecture of deep reinforcement learning when multiple actions are performed instead of a single action?

I've built a deep deterministic policy gradient reinforcement learning agent to be able to handle any games/tasks that have only one action. However, the agent seems to fail horribly when there are ...
user avatar
  • 403
5 votes
0 answers
722 views

Why isn't a target network used for the critic in on-policy actor-critic methods?

Based on my research, I've seen so many on-policy AC approaches that utilise a critic network to estimate the value function $V$. The Bellman equation for the value function is as bellow: $$ V_\pi(s_t)...
user avatar
5 votes
0 answers
72 views

Proof that there always exists a dominating policy in an MDP

I think that it is common knowledge that for any infinite horizon discounted MDP $(S, A, P, r, \gamma)$, there always exists a dominating policy $\pi$, i.e. a policy $\pi$ such that for all policies $\...
user avatar
  • 175
5 votes
0 answers
249 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 ...
user avatar
  • 51
4 votes
0 answers
75 views

What is the Thompson Sampling in simple terms?

I am looking at the different existing methods of action selection in reinforcement learning. I found several methods like epsilon-greedy, softmax, upper confidence bound and Thompson sampling. I ...
user avatar
4 votes
0 answers
87 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(\...
user avatar
  • 81
4 votes
0 answers
95 views

How to update the observation probabilities in a POMDP?

How can I update the observation probability for a POMDP (or HMM), in order to have a more accurate prediction model? The POMDP relies on observation probabilities that match an observation to a state....
user avatar
  • 65
4 votes
0 answers
73 views

Difficulty in agent's learning with increasing dimensions of continuous actions

I have been working on some RL project, where the policy is controlling the robot using its joint angles.Throughout the project I have noticed some phenomenon, which caught my attention. I have ...
user avatar
4 votes
1 answer
81 views

How can I sample the output distribution multiple times when pruning the filters with reinforcement learning?

I was reading the paper Learning to Prune Filters in Convolutional Neural Networks, which is about pruning the CNN filters using reinforcement learning (policy gradient). The paper says that the input ...
user avatar
4 votes
1 answer
148 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 ...
user avatar
4 votes
0 answers
81 views

How exactly does self-play work, and how does it relate to MCTS?

I am working towards using RL to create an AI for a two-player, hidden-information, a turn-based board game. I have just finished David Silver's RL course and Denny Britz's coding exercises, and so am ...
user avatar
4 votes
0 answers
51 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 ...
user avatar
  • 141
4 votes
0 answers
76 views

Unable to train Coach for Banana-v0 Gym environment

I have just started playing with Reinforcement learning and starting from the basics I'm trying to figure out how to solve Banana Gym with coach. Essentially ...
user avatar
4 votes
1 answer
223 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, <...
user avatar
  • 163
4 votes
1 answer
148 views

Why is my implementation of REINFORCE algorithm for portfolio optimization not converging?

I'm trying to implement the Reinforce algorithm (Monte Carlo policy gradient) in order to optimize a portfolio of 94 stocks on a daily basis (I have suitable historical data to achieve this). The idea ...
user avatar
  • 129
4 votes
0 answers
245 views

Is there a way to do reinforcement learning in POMDP?

Are there any algorithms to use reinforcement learning to learn optimal policies in partially observable Markov decision process (POMDP) i.e. when the state is not perfectly observed? More ...
user avatar
4 votes
0 answers
333 views

What is the simplest policy gradient method to implement for a problem continuous action space?

I have a problem I would like to tackle with RL, but I am not sure if it is even doable. My agent has to figure out how to fill a very large vector (let's say from 600 to 4000 in the most complex ...
user avatar
  • 145
4 votes
0 answers
626 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 ...
user avatar
  • 141
4 votes
0 answers
251 views

When using hashing in tile coding, why are memory requirements reduced and there is only a little loss of performance?

In the book "Reinforcement Learning: An Introduction" (2018) Sutton and Barto explain, on page 221, a form of tile coding using hashing, to reduce memory consumption. I have two questions ...
user avatar
  • 291
4 votes
0 answers
76 views

How define a reward function for a humanoid agent whose goal is to stand up from the ground?

I'm trying to teach a humanoid agent how to stand up after falling. The episode starts with the agent lying on the floor with its back touching the ground, and its goal is to stand up in the shortest ...
user avatar
  • 141
4 votes
0 answers
88 views

Why is there an inconsistency in the definitions of the retrace?

In Section 4.3 of the paper Learning by Playing - Solving Sparse Reward Tasks from Scratch, the authors define Retrace as $$ Q^{ret}=\sum_{j=i}^\infty\left(\gamma^{j-i}\prod_{k=i}^jc_k\right)[r(s_j,...
user avatar
  • 431
4 votes
0 answers
61 views

What research has been done on learning non-Markovian reward functions?

Recently, some work has been done planning and learning in Non-Markovian Decision Processes, that is, decision-making with temporally extended rewards. In these settings, a particular reward is ...
user avatar
4 votes
0 answers
84 views

Can deep successor representations be used with the A3C algorithm?

Deep Successor Representations(DSR) has given better performance in tasks like navigation, when compared to normal model-free RL tasks. Basically, DSR is a hybrid of model-free RL and model-based RL. ...
user avatar
4 votes
1 answer
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, ...
user avatar
4 votes
0 answers
424 views

RL to generate sentences

I want to develop a system to generate grammatically correct sentences. The input would be some words. The output would be a grammatically correct human-like sentence. Eg: Input: capital, Paris, ...
user avatar
  • 161
4 votes
1 answer
788 views

Traveling salesman problem variant: which algorithm to choose?

I have an industrial problem which I'm trying to cast as a Traveling Salesman problem (TSP) in 3D euclidian space. There are physical limitations which implies that some subpaths may or may not be ...
user avatar
  • 41
3 votes
0 answers
34 views

What is the correct interpretation of the discount factor in MDPs?

In infinite-horizon MDPs one can consider the expected discounted return from a distribution of start states as the objective[^1]. i.e. $\mathbb{E}[V^{\pi}(S_0)] = \mathbb{E}[G_0] = \mathbb{E}[\sum_{t=...
user avatar
  • 31
3 votes
1 answer
80 views

How does OpenAI-ES use Adam?

I just read that OpenAI's ES uses Adam: "OpenAI’s ES is denoted as “OptimES” (since it uses Adam optimizer)"?? I verified they are correct using the link they posted, (see es_distributed/...
user avatar
  • 296
3 votes
0 answers
99 views

What is the difference between an on-policy distribution and state visitation frequency?

On-policy distribution is defined as follows in Sutton and Barto: On the other hand, state visitation frequency is defined as follows in Trust Region Policy Optimization: $$\rho_{\pi}(s) = \sum_{t=0}^...
user avatar
3 votes
0 answers
51 views

How to prove Lemma 1.6 in the book "Reinforcement Learning: Theory and Algorithms"

I am trying to prove the following lemma from Reinforcement Learning: Theory and Algorithms on page 8. Lemma 1.6. We have that: $$ \left[(1-\gamma)\left(I-\gamma P^{\pi}\right)^{-1}\right]_{(s, a),\...
user avatar
3 votes
0 answers
57 views

How does the paper implement NEAT without a global set tracking Innovations?

I have been reading this paper on NEAT and trying to implement the algorithm in C#. For the most part, I understand everything in the paper however, there are 2 things I don't understand that confuse ...
user avatar
3 votes
0 answers
183 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 ...
user avatar
3 votes
0 answers
185 views

Why do we discount the state distribution?

In Reinforcement Learning, it is common to use a discount factor $\gamma$ to give less importance to future rewards when calculating the returns. I have also seen mention of discounted state ...
user avatar
3 votes
0 answers
163 views

How do I implement the cross-entropy-method for a RL environment with a continuous action space?

I found many tutorials and posts on how to solve RL environments with discrete action spaces using the cross entropy method (e.g., in this blog post for the OpenAI Gym frozen lake environment). ...
user avatar
  • 143
3 votes
1 answer
133 views

How can I compress the states of a reinforcement learning agent?

I'm working on a problem that involves an RL agent with very large states. These states consist of several pieces of information about the agent. The states are not images, so techniques like ...
user avatar
3 votes
0 answers
105 views

How to implement REINFORCE with eligibility traces?

The pseudocode below is taken from Barto and Sutton's "Reinforcement Learning: an introduction". It shows an actor-critic implementation with eligibility traces. My question is: if I set $\...
user avatar
3 votes
0 answers
156 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, ...
user avatar
  • 131
3 votes
0 answers
217 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 ...
user avatar
  • 117
3 votes
0 answers
281 views

In Soft Actor Critic, why is the action sampled from current policy instead of replay buffer on value function update?

While reading the original paper of Soft Actor Critic, I came across on page number 5, under equation (5) and (6) $$ J_{V}(\psi)=\mathbb{E}_{\mathbf{s}_{t} \sim \mathcal{D}}\left[\frac{1}{2}\left(V_{\...
user avatar
3 votes
1 answer
68 views

What exactly does meta-learning in reinforcement learning setting mean?

We can use DDPG to train agents to stack objects. And stacking objects can be viewed as first grasping followed by pick and place. In this context, how does meta-reinforcement learning fit? Does it ...
user avatar
  • 31
3 votes
0 answers
269 views

What is the return-to-go in reinforcement learning?

In reinforcement learning, the return is defined as some function of the rewards. For example, you can have the discounted return, where you multiply the rewards received at later time steps by ...
user avatar
  • 34.5k
3 votes
0 answers
58 views

Why is it necessary to divide the priority range according to the batch size in Prioritized Experience Replay?

According to DeepMinds's paper Prioritized Experience Replay (2016), specifically Appendix B.2.1 "Proportional prioritization" (p. 13), one should equally divide the priority range $[0, p_\...
user avatar
  • 31
3 votes
0 answers
51 views

Why weighting by lambda that sums to 1 ensures convergence in eligibility trace?

In Sutton and Barto's Book in chapter 12, they state that if weights sum to 1, then an equation's updates have "guaranteed convergence properties". Actually why it ensures convergence? There ...
user avatar
3 votes
0 answers
26 views

Mapping given probabilities to empirical probabilities

Consider following problem statement: You have given $n$ actions. You can perform any of them. Each action gives you success with some probability. The challenge is to perform given finite number of ...
user avatar
  • 211
3 votes
0 answers
46 views

Why is it the case that off-policy evaluation using importance sampling suffers from high variance?

The average return for trajectories, $V^{\pi_e}$(s) is often computed via the importance sampling estimate $$V^{\pi_e}(s) = \frac{1}{n}\sum_{i=1}^n\prod_{t=0}^{H}\frac{\pi_e(a_t | s_t)}{\pi_b(a_t|s_t)}...
user avatar
  • 1,241
3 votes
0 answers
91 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 $a=(a_1,a_2)$. ...
user avatar
  • 247
3 votes
0 answers
84 views

Convergence of a delayed policy update Q-learning

I thought about an algorithm that twists the standard Q-learning slightly, but I am not sure whether convergence to the optimal Q-value could be guaranteed. The algorithm starts with an initial ...
user avatar
3 votes
0 answers
137 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 ...
user avatar
  • 1,241

1
2 3 4 5
15