Questions tagged [rewards]

For questions related to the concept of reward, for example, in the context of reinforcement learning and Markov decision processes. For questions related to reward functions, reward design, reward shaping, reward hacking, etc., there are also those specific tags, so you use them instead of this generic one, unless your question is also about the concept of reward.

Filter by
Sorted by
Tagged with
3 votes
1 answer
24 views

$E_{\pi}[R_{t+1}|S_t=s,A_t=a] = E[R_{t+1}|S_t=s,A_t=a]$?

I would like to solve the first question of Exercise 3.19 from Sutton and Barto: Exercise 3.19 The value of an action, $q_{\pi}(s, a)$, depends on the expected next reward and the expected sum of the ...
user avatar
  • 143
1 vote
1 answer
45 views

Does it make sense to provide a DQN with negative rewards for a network with relu and sigmoid activations?

The creation of negative rewards leads to the chance of Q-values being negative. However, networks with relu or sigmoid activations, just cannot predict negative values. This will lead to a case where ...
user avatar
0 votes
0 answers
62 views

How to solve a reinforcement learning problem with a stochastic reward function?

In a discrete time system, an environment has an unknown reward probability $p(r|s,a)$. However, the transition probability $p(s'\mid s,a)$ is deterministic. In my case, the reward for the same action ...
user avatar
2 votes
0 answers
66 views

How to solve a reinforcement learning problem with changing rewards?

I'm working on a problem with non-stationary environments. The state space is discrete and limited. The action is limited too. But the reward for the same action $a$ can change. Even the reward for ...
user avatar
0 votes
0 answers
25 views

PPO: how to scale rewards

I have a custom PPO implementation and a problem that has costs rather than rewards, so I basically need to take the negative value for PPO to work. As the values are somewhat large, I've tried ...
user avatar
2 votes
1 answer
94 views

Why does the average-reward estimator for continuing tasks use the TD error?

In Sutton and Barto's RL book, section 10.3 describes how to use average reward $r(\pi)$ to define the quality of a policy, re-defining action-value function $q_\pi(s,a)$ and value function $v_\pi(s)$ ...
user avatar
  • 66
2 votes
1 answer
123 views

Why is training longer not better in reinforcement learning?

I have trained an RL agent (PPO) for 6 million steps to solve the OpenAI gym LunarLander-v2. Surprisingly, the agent performs best already after 320K steps and is getting worse after that. In the ...
user avatar
  • 213
1 vote
2 answers
93 views

What is the difference between a policy and rewards?

I don't understand the difference between a policy and rewards. Sure, a policy tells us what to do, but isn't the output of a neural network trained on rewards basically a policy (i.e. choose the ...
user avatar
1 vote
0 answers
51 views

What is commonly done for standardization/normalization of the targets in Deep Q-Learning?

I have been searching a lot about standardization/normalization of rewards and targets for the DQN algorithm. For the rewards, I now use the gym wrapper, which only scales but not shifts the rewards ...
user avatar
  • 23
1 vote
0 answers
33 views

Normalisation of reward function

Problem Currently, I have some problems defining a reward function for my RL project and mainly with how to normalise the score such that the highest possible score for all instances of the ...
user avatar
  • 21
3 votes
2 answers
545 views

What is the difference between a loss function and reward/penalty in Deep Reinforcement Learning?

In Deep Reinforcement Learning (DRL) I am having difficulties in understanding the difference between a Loss function, a reward/penalty and the integration of both in DRL. Loss function: Given an ...
user avatar
  • 131
0 votes
0 answers
54 views

How to pass the rewards in zero-sum multiplayer context when using REINFORCE?

Suppose there are two players in my zero-sum game and they play in a row like chess. And I want to learn the policy function using the REINFORCE algorithm. I have doubts about passing reward values in ...
user avatar
  • 3,099
1 vote
0 answers
32 views

Generative systems based on Schmidhuber's compression framework

In Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes ...
user avatar
  • 111
-1 votes
1 answer
158 views

Is it a bad practice to use cumulative rewards in reinforcement learning

I am using a DDPG agent for doing prediction on the position on an asset in a stock trading-like environment. I am using the cumulative reward as the reward for each timestep. Since it is trained over ...
user avatar
  • 49
1 vote
1 answer
116 views

How are rewards calculated for episodic tasks like playing chess or tic-tac-toe?

I am new to Reinforcement Learning and trying to understand the concept of reaping rewards during episodic tasks. I think in games like tic-tac-toe, rewards will be in terms of a win or lose. But does ...
user avatar
1 vote
1 answer
45 views

In the cross-entropy method, should I select state-action pairs by their immediate reward or by the episode reward?

I am trying to understand the code mechanics when selecting the elite states and elite actions. It appears clear to me that they are those that appear in the episodes with the rewards bigger than the ...
user avatar
0 votes
0 answers
142 views

How to normalize rewards in REINFORCE?

I'm trying to solve a reinforcement learning problem using a Monte Carlo policy gradient algorithm and, more specifically, REINFORCE, with rewards attributed to individual moves instead of applied to ...
user avatar
  • 121
0 votes
0 answers
21 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 ...
user avatar
2 votes
1 answer
104 views

How do I represent sample efficiency of RL rewards in mathematical notation?

I define sample efficiency as the area under the curve/graph, where $x$-axis is the number of episodes while y-axis is the cumulative reward for that episode. I would like to formally define it with a ...
user avatar
1 vote
2 answers
409 views

What is the difference between a reward and a value for a given state?

I am trying to learn reinforcement learning and I am focusing on the value iteration. I am looking at the example of grid world, and I am trying to implement it in python. While doing this, I ...
user avatar
  • 57
6 votes
1 answer
76 views

Reward interpolation between MDPs. Will an optimal policy on both ends stay optimal inside the interval?

Say I've got two Markov Decision Processes (MDPs): $$\mathcal{M_0} = (\mathcal{S}, \mathcal{A}, P, R_0),\quad\text{and}\quad\mathcal{M}_1 = (\mathcal{S}, \mathcal{A}, P, R_1)$$ Both have the same set ...
user avatar
  • 1,793
2 votes
1 answer
152 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 ...
user avatar
  • 223
0 votes
1 answer
65 views

Is is not possible to achieve average reward of more than 20-40 with simple Q-Learning

I have implemented the simple Q-Learning based solution for AI-gym's Cartpole-v0. However, despite changing hyper-parameters, and rechecking my code, I cannot get an average reward (N-running reward) ...
user avatar
  • 331
2 votes
0 answers
63 views

What is the dimensionality of these derivatives in the paper "Active Learning for Reward Estimation in Inverse Reinforcement Learning"?

I'm trying to implement in code part of the following paper: Active Learning for Reward Estimation in Inverse Reinforcement Learning. I'm specifically referring to section 2.3 of the paper. Let's ...
user avatar
2 votes
1 answer
48 views

Is there multi-agent reinforcement learning model in which (some of the) reward is given by other agent and not by the external environment?

The traditional setting of multiagent reinforcement learning (MARL) is the mode in which there is set of agents and external environment. And the reward is given to each agent - individually or ...
user avatar
  • 773
3 votes
1 answer
96 views

How do we derive the expression for average reward setting in continuing tasks?

In the average reward setting we have: $$r(\pi)\doteq \lim_{h\rightarrow\infty}\frac{1}{h}\sum_{t=1}^{h}\mathbb{E}[R_{t}|S_0,A_{0:t-1}\sim\pi]$$ $$r(\pi)\doteq \lim_{t\rightarrow\infty}\mathbb{E}[R_{t}...
user avatar
1 vote
1 answer
454 views

Why do my rewards reduce after extensive training using D3QN?

I am running a drone simulator for collision avoidance using a slight variant of D3QN. The training is usually costly (runs for at least a week) and I have observed that reward function gradually ...
user avatar
4 votes
2 answers
265 views

Why is regret so defined in MABs?

Consider a multi-armed bandit(MAB). There are $k$ arms, with reward distributions $R_i$ where $1 \leq i \leq k$. Let $\mu_i$ denote the mean of the $i^{th}$ distribution. If we run the multi-armed ...
user avatar
1 vote
0 answers
109 views

Why would DDPG with Hindsight Experience Replay not converge?

I am trying to train a DDPG agent augmented with Hindsight Experience Replay (HER) to solve the KukaGymEnv environment. The actor and critic are simple neural networks with two hidden layers (as in ...
user avatar
3 votes
1 answer
134 views

How can I fix jerky movement in a continuous action space

I am training an agent to do object avoidance. The agent has control over its steering angle and its speed. The steering angle and speed are normalized in a $[−1,1]$ range, where the sign encodes ...
user avatar
2 votes
1 answer
85 views

How do I design the rewards and penalties for an agent whose goal it is to explore a map

I am trying to train an agent to explore an unknown two-dimensional map while avoiding circular obstacles (with varying radii). The agent has control over its steering angle and its speed. The ...
user avatar
2 votes
0 answers
77 views

Should I use the discounted average reward as objective in a finite-horizon problem?

I am new to reinforcement learning, but, for a finite horizon application problem, I am considering using the average reward instead of the sum of rewards as the objective. Specifically, there are a ...
user avatar
  • 121
4 votes
2 answers
142 views

How can we prevent AGI from doing drugs?

I recently read some introductions to AI alignment, AIXI and decision theory things. As far as I understood, one of the main problems in AI alignment is how to define a utility function well, not ...
user avatar
12 votes
3 answers
856 views

Why is the reward in reinforcement learning always a scalar?

I'm reading Reinforcement Learning by Sutton & Barto, and in section 3.2 they state that the reward in a Markov decision process is always a scalar real number. At the same time, I've heard about ...
user avatar
  • 223
4 votes
3 answers
663 views

Is there an upper limit to the maximum cumulative reward in a deep reinforcement learning problem?

Is there an upper limit to the maximum cumulative reward in a deep reinforcement learning problem? For example, you want to train a DQN agent in an environment, and you want to know what the highest ...
user avatar
2 votes
1 answer
1k views

Why is the expected return in Reinforcement Learning (RL) computed as a sum of cumulative rewards?

Why is the expected return in Reinforcement Learning (RL) computed as a sum of cumulative rewards? Would it not make more sense to compute $\mathbb{E}(R \mid s, a)$ (the expected return for taking ...
user avatar
3 votes
1 answer
331 views

How is the reward in reinforcement learning different from the label in supervised learning problems?

How is the notion of immediate reward used in the reinforcement learning different from the notion of a label we find in the supervised learning problems?
user avatar
  • 35
4 votes
1 answer
638 views

Why does shifting all the rewards have a different impact on the performance of the agent?

I am new to reinforcement learning. For my application, I have found out that if my reward function contains some negative and positive values, my model does not give the optimal solution, but the ...
user avatar
1 vote
2 answers
673 views

How do I calculate the return given the discount factor and a sequence of rewards?

I know that $G_t = R_{t+1} + G_{t+1}$. Suppose $\gamma = 0.9$ and the reward sequence is $R_1 = 2$ followed by an infinite sequence of $7$s. What is the value of $G_0$? As it's infinite, how can we ...
user avatar
  • 11
1 vote
0 answers
46 views

Does importance sampling for off-policy estimation also apply to the case of negative rewards?

Importance sampling is a common method for calculating off-policy estimates in RL. I have been reading through some of the original documentation (D.G. Horvitz and D.J. Thompson, Powell, M.J. and ...
user avatar
4 votes
1 answer
888 views

Non-differentiable reward function to update a neural network

In Reinforcement Learning, when reward function is not differentiable, a policy gradient algorithm is used to update the weights of a network. In the paper Neural Architecture Search with ...
user avatar
5 votes
1 answer
553 views

If the current state is $S_t$ and the actions are chosen according to $\pi$, what is the expectation of $R_{t+1}$ in terms of $\pi$ and $p$?

I'm trying to solve exercise 3.11 from the book Sutton and Barto's book (2nd edition) Exercise 3.11 If the current state is $S_t$ , and actions are selected according to a stochastic policy $\pi$, ...
user avatar
  • 362
1 vote
1 answer
91 views

When discounted MAB is useful?

Many of multi-armed bandit(MAB) algorithms are used when the total reward is the sum of all rewards. However, in RL, the discounted reward is mainly used. Why is the discounted reward not prevailing ...
user avatar
  • 451
5 votes
2 answers
1k views

Is there any difference between reward and return in reinforcement learning?

I am reading Sutton and Barto's book on reinforcement learning. I thought that reward and return were the same things. However, in Section 5.6 of the book, 3rd line, first paragraph, it is written: ...
user avatar
  • 331
2 votes
1 answer
340 views

Can the agent wait until the end of the episode to determine the reward in SARSA?

From Sutton and Barto's book Reinforcement Learning (Adaptive Computation and Machine Learning series) (p. 99), the following definition for first-visit MC prediction, for estimating $V \sim V_\pi$ is ...
user avatar
  • 295
2 votes
1 answer
732 views

How do you manage negative rewards in policy gradients?

This old question has no definitive answer yet, that's why I am asking it here again. I also asked this same question here. If I'm doing policy gradient in Keras, using a loss of the form: ...
user avatar
  • 121
3 votes
1 answer
245 views

Why is the reward function $\text{reward} = 1/{(\text{cost}+1)^2}$ better than $\text{reward} =1/(\text{cost}+1)$?

I have implemented a simple Q-learning algorithm to minimize a cost function by setting the reward to the inverse of the cost of the action taken by the agent. The algorithm converges nicely, but ...
user avatar
  • 77
1 vote
0 answers
51 views

State-of-the-art algorithms not working on a custom RL environment

I'm trying to train a RL agent on a custom, highly stochastic environment (MDP). In order to do so I'm using existing implementations of state-of-the-art RL algorithms as provided by Stable Baselines. ...
user avatar
  • 129
1 vote
1 answer
183 views

How do you know if an agent has learnt its environment in reinforcement learning?

I'm new to reinforcement learning and trying to understand it. If you train an agent using a reinforcement learning algorithm (discrete or continuous) on an environment (real or simulated), then how ...
user avatar
3 votes
1 answer
46 views

Shouldn't expected return be calculated for some faraway time in the future $t+n$ instead of current time $t$?

I am learning RL for the first time. It may be naive, but it is a bit odd to grasp this idea that, if the goal of RL is to maximize the expected return, then shouldn't the expected return be ...
user avatar
  • 331