Questions tagged [actor-critic-methods]

For questions related to the family of reinforcement learning algorithms denoted by "actor-critic", where there is an actor (a policy) and a critic (a value function).

Filter by
Sorted by
Tagged with
30
votes
4answers
10k views

How should I handle invalid actions (when using REINFORCE)?

I want to create an AI which can play five-in-a-row/gomoku. I want to use reinforcement learning for this. I use policy gradient method, namely REINFORCE, with baseline. For the value and policy ...
13
votes
3answers
7k views

What is the difference between actor-critic and advantage actor-critic?

I'm struggling to understand the difference between actor-critic and advantage actor-critic. At least, I know they are different from asynchronous advantage actor-critic (A3C), as A3C adds an ...
7
votes
1answer
243 views

What is the purpose of the actor in actor-critic algorithms?

For discrete action spaces, what is the purpose of the actor in actor-critic algorithms? My current understanding is that the critic estimates the future reward given an action, so why not just take ...
6
votes
2answers
1k views

What is the difference between on and off-policy deterministic actor-critic?

In the paper Deterministic Policy Gradient Algorithms, I am really confused about chapter 4.1 and 4.2 which is "On and off-policy Deterministic Actor-Critic". I don't know what's the difference ...
6
votes
2answers
1k views

What are the available exploration strategies for continuous action space scenarios in RL?

I'm building a deep neural network to serve as the policy estimator in an actor-critic reinforcement learning algorithm for a continuing (not episodic) case. I'm trying to determine how to explore ...
5
votes
1answer
112 views

What are the advantages of RL with actor-critic methods over actor-only methods?

In general, what are the advantages of RL with actor-critic methods over actor-only (or policy-based) methods? This is not a comparison with the Q-learning series, but probably a method of learning ...
5
votes
1answer
218 views

Why not more TD(𝜆) in actor-critic algorithms?

Is there either an empirical or theoretical reason that actor-critic algorithms with eligibility traces have not been more fully explored? I was hoping to find a paper or implementation or both for ...
4
votes
1answer
118 views

Is reinforcement learning only about determining the value function?

I started reading some reinforcement learning literature, and it seems to me that all approaches to solving reinforcement learning problems are about finding the value function (state-value function ...
4
votes
1answer
105 views

What difference does it make whether Actor and Critic share the same network or not?

I'm learning about Actor-Critic reinforcement learning algorithms. One source I encountered mentioned that Actor and Critic can either share one network (but use different output layers) or they can ...
4
votes
1answer
119 views

What is the intuition behind the TD(0) equation with average reward, and how is it derived?

In chapter 10 of Sutton and Barto's book (2nd edition) is given the equation for TD(0) error with average reward (equation 10.10): $$\delta_t = R_{t+1} - \bar{R} + \hat{v}(S_{t+1}, \mathbf{w}) - \hat{...
4
votes
1answer
180 views

Would you categorize policy iteration as an actor-critic reinforcement learning approach?

One way of understanding the difference between value function approaches, policy approaches and actor-critic approaches in reinforcement learning is the following: A critic explicitly models a value ...
4
votes
2answers
210 views

How is parallelism implemented in RL algorithms like PPO?

There are multiple ways to implement parallelism in reinforcement learning. One is to use parallel workers running in their own environments to collect data in parallel, instead of using replay memory ...
4
votes
1answer
185 views

What is the advantage of using more than one environment with the advantage actor-critic?

make_env = lambda: ptan.common.wrappers.wrap_dqn(gym.make("PongNoFrameskip-v4")) envs = [make_env() for _ in range(NUM_ENVS)] Here is a code you can look at. ...
4
votes
1answer
118 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, <...
4
votes
1answer
253 views

Understanding multi-iteration updates of the model in the Proximal Policy Optimization algorithm

I have a general question about the updating of the network/model in the PPO algorithm. If I understand it correctly, there are multiple iterations of weight updates done on the model with data that ...
3
votes
1answer
177 views

Where does entropy enter in Soft Actor-Critic?

I am currently trying to understand SAC (Soft Actor-Critic), and I am thinking of it as a basic actor-critic with the entropy included. However, I expected the entropy to appear in the Q-function. ...
3
votes
1answer
494 views

Meaning of Actor Output in Actor Critic Reinforcement Learning

In actor critic, The equations for calculating the loss in actor critic are an actor loss (parameterized by $\theta$) $$log[\pi_\theta(s_t,a_t)]Q_w(s_t,a_t)$$ and a critic loss (parameterized by ...
3
votes
1answer
173 views

How does being on-policy prevent us from using the replay buffer with the policy gradients?

One of the approaches to improving the stability of the Policy Gradient family of methods is to use multiple environments in parallel. The reason behind this is the fundamental problem we discussed in ...
3
votes
1answer
1k views

Why can't my implementation of the Actor-Critic algorithm achieve good results in the 2048 game?

I implemented the Actor-Critic with n-step TD prediction to learn to play the 2048 game For the environment, I don't use this 2048 implementation. I use a simple one without any graphical interface, ...
3
votes
0answers
57 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 ...
3
votes
0answers
57 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 $\...
3
votes
0answers
180 views

Should noise (such as OU) be decreased over time in actor / critic algorithms?

In most of RL algorithms I saw, there is a coefficient that reduces actions exploration over time, to help convergence. But in Actor-Critic, or other algorithms (A3C, DDPG, ...) used in continuous ...
3
votes
1answer
115 views

Can A3C update the policy / critic on a local machine without needing to copy?

To make A2C into A3C you make it asynchronous. From what I understand the 'correct' way to do that is to thread off workers with a copy of the policy and critic, and then return the state/action/...
2
votes
1answer
353 views

Can I apply DQN or policy gradient algorithms in the contextual bandit setting?

I have a problem which I believe can be described as a contextual bandit. More specifically, in each round, I observe a context from the environment consisting of five continuous features, and, ...
2
votes
2answers
185 views

Advantage computed the wrong way?

Here is the code written by Maxim Lapan. I am reading his book (Deep Reinforcement Learning Hands-on). I have seen a line in his code which is really weird. In the accumulation of the policy gradient $...
2
votes
2answers
208 views

What is the gradient of the objective function in the Soft Actor-Critic paper?

In the paper Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor, they define the loss function for the policy network as $$ J_\pi(\phi)=\mathbb E_{s_t\...
2
votes
1answer
609 views

How do I calculate the policy in the Proximal Policy Optimization algorithm?

I recently watched the video on Proximal Policy Optimization (PPO). Now, I want to upgrade my actor-critic algorithm written in PyTorch with PPO, but I'am not sure how the new parameters / thetas are ...
2
votes
1answer
529 views

Why isn't my implementation of A2C for the the atari pong game converging?

I have two different implementations with PyTorch of the Atari Pong game using A2C algorithm. Both implementations are similar, but some portion are different. https://colab.research.google.com/drive/...
2
votes
1answer
72 views

Learning policy where action involves discrete and continuous parameters

Typically it seems like reinforcement learning involves learning over either a discrete or a continuous action space. An example might be choosing from a set of pre-defined game actions in Gym Retro ...
2
votes
1answer
53 views

Why a single trajectory can be used to update the policy network $\theta$ in A3C?

The Deep RL bootcamp on policy gradient techniques gives the update equation for the policy network in A3C as $\theta_{i+1} = \theta_i + \alpha \times 1/m \sum_{k=1}^m\sum_{t=0}^{H-1}\nabla_{\theta}...
2
votes
1answer
283 views

What is the difference between A2C and running an agent in an environment vector?

I've implemented A2C. I'm now wondering why would we have multiple actors walk around the environment and gather rewards, why not just have a single agent run in an environment vector? I personally ...
2
votes
1answer
133 views

Why is gradient ascent necessary when training Actor Critic agents?

I have read a lot on Actor Critic and I'm not convinced that there is a qualitative difference doing direct gradient updates on the network and slightly adjusting a soft-max output in the direction of ...
2
votes
1answer
479 views

Are A2C or A3C suitable for episodic tasks where the reward is delivered only at the end of the episode?

My understanding of the main idea behind A2C / A3C is that we run small segments of an episode to estimate the return using a trainable value function to compensate for the unseen final steps of the ...
2
votes
1answer
449 views

In online one step actor critic, why does the weights update become less significant as the episode progresses?

The Reinforcement Learning Book by Richard Sutton et al, section 13.5 shows an online actor critic algorithm. Why do the weights updates depend on the discount factor via $I$? It seems that the more ...
2
votes
1answer
150 views

Why is the actor-critic algorithm limited to using on-policy data?

Why is the actor-critic algorithm limited to using on-policy data? Or can we use the actor-critic algorithm with off-policy data?
2
votes
0answers
44 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} ...
2
votes
0answers
87 views

Original source of the TD Advantage Actor-Critic algorithm?

What is the original source of the TD Advantage Actor-Critic algorithm? I found the following tutorial really helpful for learning the algorithm: https://medium.com/@asteinbach/actor-critic-using-...
2
votes
0answers
72 views

What is the gradient of the Q function with respect to the policy's parameters?

I have been recently studying Actor-Critic algorithms, and I ran into the following question. Let $Q_{\omega}$ be the critic network, and $\pi_{\theta}$ be the actor. It is known that in order to ...
2
votes
1answer
245 views

How to set the target for the actor in A2C?

I did a simple Actor-Critic implementation in Keras using 2 networks where the critic learns the Q-Values of every action, and the actor predicts probabilities for choosing each action. In training, ...
2
votes
0answers
137 views

How does the update rule for the one-step actor-critic method work?

Can you please elucidate the math behind the update rule for the critic? I've seen in other places that just a squared distance of $R + \hat{v}(S', w) - \hat{v}(S,w)$ is used, but Sutton suggests an ...
1
vote
1answer
582 views

How can the derivative of a neural network be calculated, given no mathematical expression?

Neural networks (NNs) are used as approximators in reinforcement learning (RL). To update the policy in RL, the actor network's gradients w.r.t its weights are needed. Since NN doesn't have a ...
1
vote
1answer
667 views

How is the actor-critic algorithm guaranteed to converge?

From my understanding, the critic evaluates the policy (actor) following dynamic programming (DP) or approximate dynamic programming (ADP) scheme, which should converge to the optimal value function ...
1
vote
1answer
98 views

What does the notation $\partial \theta_{\pi}$ mean in this actor-critic update rule?

One of the steps in the actor-critic algorithm is $$\partial \theta_{\pi} \gets \partial \theta_{\pi} + \nabla_{\theta}\log\pi_{\theta} (a_i | s_i) (R - V_{\theta}(s_i))$$ For me, $\theta$ are just ...
1
vote
1answer
276 views

A2C Critic Loss Interpretation

I'm working on an Advantage A2C implementation, and I just finished creating the value network $\hat{V_{\phi}}$. I train this network with the standard MSE loss of discounted rewards-to-go:$$\|\hat{V_{...
1
vote
1answer
120 views

What is the most biologically plausible representation for the actor and critic?

Which representation is most biologically plausible for actor nodes? For example, actions represented across several output nodes which may be either mutually exclusive with each other (e.g., go ...
1
vote
1answer
131 views

Can neuro-evolution methods be combined with A3C?

As a amateur researcher and tinkerer, I've been reading up on neuro-evolution networks (e.g. NEAT) as well as the A3C RL approach presented by Mnih et al and got to wondering if anyone has ...
1
vote
0answers
11 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)...
1
vote
0answers
22 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
1answer
79 views

Advantage Actor Critic model implementation with Tensorflowjs

I am trying to implement an Actor Critic method that controls an RC car. For this I have implemented a simulated environment and actor critic tensorflowjs models. My intention is to train a model to ...