Questions tagged [policy-gradients]

For questions related to reinforcement learning algorithms often referred to as "policy gradients" (or "policy gradient algorithms"), which attempt to directly optimise a parameterised policy (without first attempting to estimate value functions) using gradients of an objective function with respect to the policy's parameters.

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

What specifically is the gradient of the log of the probability in policy gradient methods (reinforcement learning)

I am getting tripped up slightly by how specifically the gradient is calculated in policy gradient methods (just the intuitive understanding of it). https://math.stackexchange.com/questions/2845971/...
0
votes
0answers
21 views

FrozenLake-v0 not training using REINFORCE

I am implementing a simple REINFORCE (policy gradient) algorithm for openAI's FrozenLake-v0 environment. However, it does not seem to learn anything at all. I have used the same neural architecture ...
0
votes
0answers
15 views

How to fix loss?

I am trying to use the Vanilla Policy Gradient algorithm to solve MountainCar-v0. I have taken the code from pytorch repo. ...
4
votes
1answer
65 views

Why exactly was previously believed that the deterministic policy gradient did not exist?

I'm reading the paper Deterministic Policy Gradient Algorithms, David Silver et al. First of all, in the introduction, the author says that It was previously believed that the deterministic policy ...
1
vote
0answers
33 views

Why does this PPO agent doesn't learn at all?

I am trying to code a Proximal Policy Optimization algorithm in Pytorch by myself for the OpenAI gym preduum-v0 environment. However, I find my learning curve is ...
1
vote
0answers
21 views

Can we learn a policy network via a sequence of manually determined actions?

In policy gradients, is it possible to learn the policy if the chain of actions is selected and performed manually/externally (e.g. by myself or by someone else who I have no influence over)? For ...
0
votes
1answer
44 views

Can a Reinforcement Learning problem with multiple simultaneous actions be formalized as a Multiagent Partially Observable Markov Decision Process?

Consider the following decision making problem. We have a controller that selects locations from a grid of coordinates and captures an image (observation $o_t$) with a camera at each location (action $...
0
votes
0answers
18 views

multi agent deep deterministic policy gradient for discrete actions

I am solving a multi agent problem where each agent has a critic and actor. The problem I am solving has discrete actions and discrete states. I came cross multi-agent deep deterministic policy ...
0
votes
0answers
21 views

DDPG or PPO don't work well with my custom non gym environment

I have a project to control a robot with right and left wheel speeds, and my step time is not constant. Because my outputs are continuous (right wheel speed, left wheel speed, and time step), I try to ...
1
vote
1answer
36 views

How to interpret the policy gradient expression in reinforcement learning?

I'm currently going through the OpenAI's spinning up introduction course to reinforcement learning. On one of the final sections, they derive an expression for the gradient of the undiscounted return ...
2
votes
1answer
36 views

Why adding a baseline doesn't affect the policy gradient?

On the OpenAI's Spinning Up, they justify the fact that adding a baseline $b(s_t)$ in the policy gradient doesn't change its gradient by saying that this is an immediate consequence of the EGLP Lemma ...
0
votes
0answers
19 views

Multi-agent policy gradient, 1 total reward instead of reward in each step, 2 changing action space

I am new in reinforcement learning and not sure I have the right understanding of multi-agent policy gradient. 1, in my question, each agent has its own action space. When doing the sampling, for each ...
1
vote
1answer
72 views

How does the neural network learn when used in the REINFORCE algorithm?

As per my understanding, you run an entire episode, which contains many steps, and then back-propagate using just 1 loss value. How does the neural network learn to differentiate between good and bad ...
0
votes
0answers
16 views

A2C value function outputs keep increasing

I was implementing the A2C algorithm with as close to baseline setup as possible, and this is the code I came up with. The problem is that even after multiple rechecks, the program isn't showing ...
0
votes
0answers
12 views

How do I quantify the difference in sample efficiency for two almost similar methods?

I am comparing my coded TD3 (Twin-Delayed DDPG) and the same TD3 (same hyperparameters) but with Priority Replay Buffer instead of a normal Replay Buffer. From what I have read, PER (Priority ...
0
votes
1answer
44 views

How to formulate discounted return in cartpole?

I am trying to formulate a problem that aims to prolong the lifetime of the simulation, the same as the Cartpole problem. I aware that there are two types of return: finite horizon undiscounted ...
4
votes
1answer
483 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
93 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
90 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{...
1
vote
0answers
216 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 - <...
1
vote
1answer
51 views

Is the policy gradient expression in Fundamentals of Deep Learning wrong?

I don't understand the policy gradient as explained in Chapter-9 (Deep Reinforcement Learning) of the book Fundamentals of deep learning. Here is the whole paragraph: Policy Learning via Policy ...
1
vote
0answers
22 views

Is there any research on the application of policy gradients to problems where the selection of an action requires the selection of another one?

I am working on a problem and want to explore if it can be solved with PPO (or other policy gradient methods). The problem is that the action space is a bit special, compared to classic RL ...
1
vote
1answer
104 views

Why does the implementation of REINFORCE algorithm minimize the gradient term but not the loss?

I read the book "Foundation of Deep Reinforcement Learning, Laura Graesser and Wah Loon Keng", and when I go through the REINFORCE algorithm, they show the objective function: $$ J\left(\...
1
vote
0answers
21 views

Understanding advantage estimator in proximal policy optimization

I was reading Proximal Policy Optimization paper. It states following: The advantage estimator used is: $\hat{A}_t=-V(s_t)+r_t+\gamma r_{t+1}+...+\gamma^{T-t+1}r_{T-1}+\color{blue}{\gamma^{T-t}}V(s_T)...
0
votes
0answers
26 views

Which off-policy policy gradient estimator has lower variance?

Let $\pi_\theta$ be a target policy and $\beta_\theta$ be a behavior policy. I have seen the following 2 policy gradient estimators in the literature: $$ \operatorname*{E}_{\tau \sim \beta_\theta} \...
1
vote
1answer
23 views

How to have zero value or a value between 200 and 400 in the output of a deep learning model?

I want to implement a DDPG method and obviously, the action space will be continuous. I have three outputs. The first output should be zero or a value between 200 and 400, and the other outputs have ...
1
vote
1answer
55 views

Why is the behaviour policy denoted by $\beta$ and the exploration policy by $ \mu'$ in the DDPG paper?

I am learning about the deep deterministic policy gradient (DDPG) (Lillicrap et al, 2016) and got confused about the notation of the behavior policy. Lillicrap et al. denote the policy gradient by $$\...
1
vote
1answer
93 views

It is mathematically correct to use a Policy Gradient method for 1-step trajectories?

I have come across a Google paper that uses the REINFORCE algorithm (a Policy Gradient Method) for a case where the trajectory of the episodes it proposes would be ...
1
vote
0answers
57 views

Confusion about computing policy gradient with automatic differentiation ( material from Berkeley CS285)

I am taking Berkeley’s CS285 via self-study. On this particular lecture regarding Policy Gradient, I am very confused about the inconsistency between the concept explanation and the demonstration of ...
1
vote
0answers
38 views

Understanding loss function gradient in asynchronous advantage actor-critic (A3C) algorithm

This is a question I posted here. I am asking it on this StackExchange branch as well, so that more people who could potentially answer get to see the question. In the A3C algorithm from the original ...
1
vote
0answers
38 views

Understanding policy gradient derivation with importance sampling

I was watching a lecture on Policy Gradient and had difficulty following it when importance sampling was introduced. It was shown that the gradient of the objective can be written as $$\nabla_\theta U(...
3
votes
0answers
70 views

$\nabla \log \pi$ with respect to some parameters constantly being zero

I am new to reinforcement learning. May I ask a simple (and maybe a bit silly) question here? I am trying to use the "one-step actor-critic" method to train a robot on a gridworld. Let's ...
1
vote
0answers
25 views

Why do we use big batch/epoch size in policy gradient methods (vpg specifically)?

I am re-implementing vpg and using Spinning Up as reference implementation. I noticed that the default epoch size is 4000. I also see cues in papers that big batch size is quite standard. My ...
1
vote
0answers
220 views

Designing Policy-Network for Deep-RL with Large, Variable Action Space

I am attempting a project involving training an agent to play a game using deep reinforcement learning. This project has a few features that complicate the design of the neural network: The action ...
2
votes
0answers
97 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
0answers
99 views

Why does PPO lead to a worse performance than TRPO in the same task?

I am training an agent with an Actor-Critic network and update it with TRPO so far. Now, I tried out PPO and the results are drastically different and bad. I only changed from TRPO to PPO, the rest of ...
3
votes
0answers
24 views

Understanding neural network achitectures in policy gradient reinforcement learning for continuous state and action space

I am trying to train a neural network using reinforcement learning / policy gradient methods. The states, i.e. the inputs, as well as the actions I am trying to sample are vectors with each element ...
3
votes
0answers
182 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_{\...
2
votes
0answers
50 views

Problems with gradient-biased actor critic methods

To my knowledge, there are at least 6 different variants of Actor Critic: \begin{array}{l l l l} \text{actor gradient} & \text{critic gradient} & \text{actor gradient biased} & \text{name} ...
1
vote
0answers
47 views

What are the disadvantages of actor-only methods with respect to value-based ones?

While the advantages of actor-only algorithms, the ones that search directly the policy without the use of the value function, are clear (possibility of having a continuous action space, a stochastic ...
5
votes
1answer
656 views

What is the loss for policy gradients with continuous actions?

I know with policy gradients used in an environment with a discrete action space are updated with $$ \Delta \theta_{t}=\alpha \nabla_{\theta} \log \pi_{\theta}\left(a_{t} \mid s_{t}\right) v_{t} $$ ...
7
votes
1answer
266 views

What, exactly, does the REINFORCE update equation mean?

I understand that this is the update for the parameters of a policy in REINFORCE: $$ \Delta \theta_{t}=\alpha \nabla_{\theta} \log \pi_{\theta}\left(a_{t} \mid s_{t}\right) v_{t} $$ Where 𝑣𝑡 is ...
3
votes
2answers
80 views

Why does (not) the distribution of states depend on the policy parameters that induce it?

I came across the following proof of what's commonly referred to as the log-derivative trick in policy-gradient algorithms, and I have a question - While transitioning from the first line to the ...
2
votes
0answers
34 views

What's an example of a simple policy but a complex value function?

Hado van Hasselt, a researcher at DeepMind, mentioned in one of his videos (from 7:20 to 8:20) on Youtube (about policy gradient methods) that there are cases when the policy is very simple compared ...
2
votes
1answer
154 views

Generation of 'new log probabilities' in continuous action space PPO

I have a conceptual question for you all that hopefully I can convey clearly. I am building an RL agent in Keras using continuous PPO to control a laser attached to a pan/tilt turret for target ...
4
votes
1answer
1k 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
131 views

Why does REINFORCE work at all?

Here's a screenshot of the popular policy-gradient algorithm from Sutton and Barto's book - I understand the mathematical derivation of the update rule - but I'm not able to build intuition as to why ...
1
vote
0answers
30 views

Customized food for persons based on their profile using Reinforcement learning

I am newbie to Reinforcement Learning, this is my idea - Agent(food provider) has to select a food based on the environment(based on the user profile). Here the reward will be given to the agent based ...
1
vote
1answer
516 views

DDPG doesn't converge for MountainCarContinuous-v0 gym environment

I am trying to implement Deep Deterministic policy gradient algorithm by referring to the paper Continuous Control using Deep Reinforcement Learning on the MountainCarContinuous-v0 gym environment. I ...
5
votes
1answer
134 views

Comparing the derivation of the Deterministic Policy Gradient Theorem with the standard Policy Gradient Theorem

I would like to understand the difference between the standard policy gradient theorem and the deterministic policy gradient theorem. These two theorem are quite different, although the only ...