Questions tagged [policies]

For questions related to policies (as defined in reinforcement learning or other AI sub-fields).

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

Is Q-learning only capable of learning a deterministic policy?

I was following a reinforcement learning course on coursera and in this video at 2:57 the instructor says Expected SARSA and SARSA both allow us to learn an optimal $\epsilon$-soft policy, but, Q-...
user avatar
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
0 votes
1 answer
38 views

What's the benefit of repeating an action for a consecutive number of time steps?

What's the benefit of repeating an action for a consecutive number of time steps? Is there a way to tell if an agent in a given environment might perform better from repeated actions? I came across an ...
user avatar
  • 509
2 votes
1 answer
106 views

Is the initialisation of $V(s)$ and $\pi(s)$ really arbitrary in policy iteration?

In Sutton and Barto's book (Reinforcement learning: An introduction. MIT press, 2018), the algorithm "Policy Iteration" is: Here, $V(s)$ is initialized arbitrarily, meaning that I can ...
user avatar
  • 121
3 votes
0 answers
74 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
1 answer
121 views

Can an optimal policy have a value function that has a smaller value for a state than a non-optimal policy?

I'm starting to learn about the Bellman Equation and a question came to my mind. A policy $\pi$ is optimal if the value $v_\pi(s)$ is greater or equal than the value $v_{\pi'}(s)$ for all states $s \...
user avatar
3 votes
1 answer
75 views

In reinforcement learning, why are policies defined as functions of states and not observations?

I am new to RL and I am following Sutton & Barto's book. My doubt is, when we talk about the policy of our agent, we say it is the probability of taking some action $a$ given the state $s$. ...
user avatar
2 votes
1 answer
71 views

Why must the value of a state under an optimal policy equal the expected return for the best action from that state?

The Sutton and Barto reinforcement learning textbook states that the value of a state under an optimal policy must equal the expected return for the best action from that state. That is, $$v_*(s) = \...
user avatar
1 vote
0 answers
33 views

How to allow RL systems to find better policies after code changes?

Suppose that in version 1 of a reinforcement-learning system an optimal policy $A$ got generated for executing a task. But, in a newer version 2 of that application (with new code changes), there ...
user avatar
5 votes
0 answers
68 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
0 votes
0 answers
89 views

DQN learns to always choose the same action for all states

I have created an RL model that uses QBased policy with a neural network for estimating Q values. My action space is of 27 actions, where each action is a 3 tuple where each value can be 1, 2 or 3. ...
user avatar
0 votes
1 answer
166 views

Are optimal policies always deterministic, or can there also be optimal policies that are stochastic?

Let $M$ be an MDP with two states, $A$ and $B$, where $A$ is the starting state, and you always transit to the final state $B$ using two possible actions. $A_1$ gives you rewards that are normally ...
user avatar
  • 35
1 vote
1 answer
61 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 ...
user avatar
3 votes
1 answer
224 views

How can we find the value function by solving a system of linear equations without knowing the policy?

An MDP is a Markov Reward Process with decisions, it’s an environment in which all states are Markov. This is what we want to solve. An MDP is a tuple $(S, A, P, R, \gamma)$, where $S$ is our state ...
user avatar
  • 35
2 votes
1 answer
75 views

Why do I get the best policy before Q values converge using DQN?

I have implemented DQN algorithm and wonder why during testing, the best performance is achieved by a policy from about 300 episode, when mean Q values converge at about 800 episode? Mean Q-values ...
user avatar
2 votes
1 answer
59 views

Is a learned policy, for a deterministic problem, trained in a supervised process, a stochastic policy?

If I trained a neural network with 4 outputs (one for each action: move down, up, left, and right) to move an agent through a grid (deterministic problem). The output of the neural network is a ...
user avatar
  • 21
2 votes
1 answer
425 views

What is a "learned policy" in Q-learning?

I am completing an assignment at the moment. One of the assignment questions asks how you identified the learned policy and how you obtained it. The question is a reinforcement learning question, and ...
user avatar
2 votes
0 answers
34 views

Does there necessarily exist "dominated actions" in a MDP?

In a Markov Decision Process, is it possible that there exists no "dominated action"? I define a dominated action the following way: we say that $(s,a)$ is a dominated action, if $\forall \...
user avatar
  • 31
1 vote
0 answers
16 views

How to obtain the policy in the form of a finite-state controller from the value function vectors over the belief space of the POMDP?

I was reading this paper by Hansen. It says the following: A correspondence between vectors and one-step policy choices plays an important role in this interpretation of a policy. Each vector in $\...
user avatar
  • 211
1 vote
0 answers
87 views

What kind of reinforcement learning method does AlphaGo Deepmind use to beat the best human Go player?

In reinforcement learning, there are model-based versus model-free methods. Within model-based ones, there are policy-based and value-based methods. AlphaGo Deepmind RL model has beaten the best Go ...
user avatar
3 votes
1 answer
233 views

Are there some notions of distance between two policies?

I want to determine some distance between two policies $\pi_1 (a \mid s)$ and $\pi_2 (a \mid s)$, i.e. something like $\vert \vert \pi_1 (a \mid s) - \pi_2(a \mid s) \vert \vert$, where $\pi_i (a\mid ...
user avatar
  • 287
0 votes
1 answer
81 views

How to use and update a shared/global policy between Reinforcement Learning Agents

I would be grateful for some guidance on a RL problem I am trying to solve where multiple RL agents use a common/global policy at the initial state of an episode in the RL Environment, and then update ...
user avatar
  • 15
5 votes
2 answers
417 views

Given two optimal policies, is an affine combination of them also optimal?

If there are two different optimal policies $\pi_1, \pi_2$ in a reinforcement learning task, will the linear combination (or affine combination) of the two policies $\alpha \pi_1 + \beta \pi_2, \alpha ...
user avatar
0 votes
0 answers
48 views

Off-policy full-random training in easy-to-explore environment

Let say we are in an environment where a random agent can easily explore all the states of an environment (for example: tic-tac-toe). In those environments, using off-policy algorithm, is it a good ...
user avatar
  • 266
4 votes
1 answer
210 views

Why do we have two similar action selection strategies for UCB1?

In the literature, there are at least two action selection strategies associated with the UCB1's action selection strategy/policy. For example, in the paper Algorithms for the multi-armed bandit ...
user avatar
  • 33.8k
1 vote
0 answers
42 views

Dynamically adapting activation function

I am training a network through reinforcement learning. The policy network learns rotations, but depending on the actual input (state), the output of the network should be restricted to be in certain ...
user avatar
  • 31
2 votes
0 answers
51 views

What are some other real-life examples of simple policies but complex value functions?

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 ...
user avatar
3 votes
1 answer
79 views

Why doesn't value iteration use $\pi(a \mid s)$ while policy evaluation does?

I was looking at the Bellman equation, and I noticed a difference between the equations used in policy evaluation and value iteration. In policy evaluation, there was the presence of $\pi(a \mid s)$, ...
user avatar
2 votes
0 answers
26 views

Do we assume the policy to be deterministic when proving the optimality?

In reinforcement learning, when we talk about the principle of optimality, do we assume the policy to be deterministic?
user avatar
1 vote
2 answers
223 views

Why is the optimal policy for an infinite horizon MDP deterministic?

Could someone please help me gain some intuition as to why the optimal policy for a Markov Decision Process in the infinite horizon case (agent acts forever) is deterministic?
user avatar
2 votes
1 answer
71 views

Is a policy in reinforcement learning analogous to a field such as APF?

If a policy maps states to actions in reinforcement learning, then for a path planning with obstacles, can't we simply use Artificial Potential Field fields for path planning and model policy ...
user avatar
  • 171
2 votes
1 answer
78 views

Is it common to have extreme policy's probabilities?

I have implemented several policy gradient algorithms (REINFORCE, A2C, and PPO) and am finding that the resultant policy's action probability distributions can be rather extreme. As a note, I have ...
user avatar
3 votes
1 answer
134 views

What is meant by "generate the data" in describing the difference between on-policy and off-policy?

From the book: Sutton, Richard S.,Barto, Andrew G.. Reinforcement Learning (Adaptive Computation and Machine Learning series) (p. 100). The MIT Press. Kindle Edition. " following is stated: "...
user avatar
  • 295
1 vote
1 answer
123 views

Do we need multiple parallel environments to train in batches an on-policy algorithm?

When using an on-policy method in reinforcement learning, like advantage actor-critic, you shouldn't use old data from an experience buffer, since a new policy requires new data. Does this mean that ...
user avatar
  • 111
1 vote
0 answers
88 views

Does the off-policy evaluation work for non-stationary policies?

As the title says, in reinforcement learning, does the off-policy evaluation work for non-stationary policies? For example, IS (importance sampling)-based estimators, such as weighted IS or doubly ...
user avatar
  • 207
2 votes
1 answer
146 views

Why is there an inconsistency between my calculations of Policy Iteration and this Sutton & Barto's diagram?

In equation 4.9 of Sutton and Barto's book on page 79, we have (for the policy iteration algorithm): $$\pi'(s) = arg \max_{a}\sum_{s',r}p(s',r|s,a)[r+\gamma v_{\pi}(s')]$$ where $\pi$ is the previous ...
user avatar
2 votes
2 answers
251 views

Why is the policy not a part of the MDP definition?

I'm reading an article on reinforcement learning, and I don't understand why the agent's policy $\pi$ is not part of definition of Markov Decision process(MDP): Bu, Lucian, Robert Babu, and Bart De ...
user avatar
3 votes
1 answer
554 views

Why do RL implementations converge on one action?

I have seen this happening in implementations of state-of-the-art RL algorithms where the model converges to a single action over time after multiple training iterations. Are there some general ...
user avatar
3 votes
2 answers
145 views

How does the AlphaGo Zero policy decide what move to execute?

I was going through the AlphaGo Zero paper and I was trying to understand everything, but I just can't figure out this one formula: $$ \pi(a \mid s_0) = \frac{N(s_0, a)^{\frac{1}{\tau}}}{\sum_b N(s_0,...
user avatar
  • 33
2 votes
3 answers
93 views

How to estimate a behavior policy for off-policy learning based on data?

I have a dataset which includes states, actions, and reward. The dataset includes information on the transition, i.e., $p(r,s' \mid s,a)$. Is there a way to estimate a behavior policy from this ...
user avatar
3 votes
2 answers
1k views

Why do value iteration and policy iteration obtain similar policies even though they have different value functions?

I am trying to implement value and policy iteration algorithms. My value function from policy iteration looks vastly different from the values from value iteration, but the policy obtained from both ...
user avatar
2 votes
1 answer
106 views

Why do I need an initial arbitrary policy to implement value iteration algorithm

I've been recently given an assignment based on Reinforcement Learning and I'm supposed to implement the value iteration algorithm in a grid environment. The assignment: My doubt is why do I even ...
user avatar
  • 121
2 votes
2 answers
442 views

Is my understanding of the value function, Q function, policy, reward and return correct?

I'm a beginner in the RL field, and I would like to check that my understanding of certain RL concepts. Value function: How good it is to be in a state S following policy π. ...
user avatar
  • 113
1 vote
2 answers
147 views

Why does the policy $\pi$ affect the Q value?

From my understanding, the policy $\pi$ is basically how the agent acts (i.e. the actions it will take in each state). However, I am confused about the Q value and how it is "affected" by a policy. ...
user avatar
  • 113
3 votes
0 answers
74 views

Representation of state space, action space and reward system for Reinforcement Learning problem

I am trying to solve the problem of an agent dynamically discovering(start with no information about the environment) the environment and to explore as much of the environment as possible without ...
user avatar
0 votes
0 answers
89 views

How to understand and visualize a trained RL agent's policy when the state space is high dimensional?

What are typical ways to understand and visualize a trained RL agent's policy when the state space is of high dimension (but not images)? For example, suppose state and action are denoted by $s=(...
user avatar
  • 121
3 votes
3 answers
446 views

Is the policy really invariant under affine transformations of the reward function?

In the context of a Markov decision process, this paper says it is well-known that the optimal policy is invariant to positive affine transformation of the reward function On the other hand, ...
user avatar
  • 171
2 votes
0 answers
55 views

Can the importance sampling estimator have a non-stationary behaviour policy even if the target policy is stationary?

The inverse propensity score (IPS) estimator, which is used for off-policy evaluation in a contextual bandit problem, is well explained in the paper Doubly Robust Policy Evaluation and Optimization. ...
user avatar
  • 207
2 votes
0 answers
128 views

AlphaGo Zero: Does the policy head give a probability for every possible move?

If I understood correctly, the AlphaGo Zero network returns two values: a vector of logit probabilities p and a value v. My question is: in this vector that it is outputted, do we have a probability ...
user avatar
4 votes
1 answer
464 views

Why does having a fixed policy change a Markov Decision Process to a Markov Reward Process?

If a policy is fixed, it is said that a Markov Decision Process (MDP) becomes a Markov Reward Process (MRP). Why is this so? Aren't the transitions and rewards still parameterized by the action and ...
user avatar
  • 43