Questions tagged [proofs]

For questions that ask about or call for proofs for specific assertions, whether they be proofs of theorems or corollaries, proofs of concept through working implementation, counter proofs, or counter examples.

Filter by
Sorted by
Tagged with
1
vote
0answers
46 views

Can the law of iterated expectation be used on the inner expectation of the DQN cost function described in the DQN paper

Is the expression for the DQN cost function, Equation (2) of the DQN paper $$\begin{align}L_1 &= E_{\mu,\pi}\left[\left(y_i - q(s,a;\theta)\right)^2\right]\\ &=E_{\mu,\pi}\left[\left(E_{\...
1
vote
1answer
63 views

If $h_1(n)$ is admissible, why does A* tree search with $h_2(n) = 3h_1(n)$ return a path that is at most thrice as long as the optimal path?

Consider a heuristic function $h_2(n) = 3h_1(n)$. Where $h_1(n)$ is admissible. Why are the following statements true? $A^*$ tree search with $h_2(n)$ will return a path that is at most thrice as ...
5
votes
2answers
177 views

Why is “-0.5 * torch.sum(1 + sigma - mu.pow(2) - sigma.exp())” in Pytorch equivalent to the KL?

I found this code for the loss function of a VAE: -0.5 * torch.sum(1 + sigma - mu.pow(2) - sigma.exp()) From this link: https://debuggercafe.com/getting-started-...
3
votes
1answer
115 views

Is my proof of equation 0.6 in the book “Reinforcement Learning: Theory and Algorithms” correct?

In Sham Kakade's Reinforcement Learning: Theory and Algorithms, this equation (page 17) is used preceding the proof of performance difference lemma. I am attempting to prove equation 0.6. Here is my ...
4
votes
0answers
65 views

When do two identical neural networks have uncorrelated errors?

In Chapter 9, section 9.1.6, Raul Rojas describes how committees of networks can reduce the prediction error by training N identical neural networks and averaging the results. If $f_i$ are the ...
2
votes
1answer
90 views

How to prove the formula of eligibility traces operator in reinforcement learning?

I don't understand how the formula in the red circle is derived. The screenshot is taken from this paper
2
votes
0answers
21 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 \...
1
vote
1answer
28 views

How do I prove that the MSE is zero when all predictions are equal to the corresponding labels?

In the back-propogation algorithm, the error term is: $$ E=\frac{1}{2}\sum_k(\hat{y}_k - y_k)^2, $$ where $\hat{y}_k$ is a vector of outputs from the network, $y_k$ is the vector of correct labels (...
0
votes
0answers
17 views

Is the PR AUC invariant under label flip?

The ROC-AUC curve is invariant under a flip of the labels. I don't know if it's a famous result, so I will give the proof below. My question is if the PR-AUC curve also has this property. I have not ...
1
vote
1answer
206 views

Is there a proof to explain why XOR cannot be linearly separable?

Can someone explain to me with a proof or example why you can't linearly separate XOR (and therefore need a neural network, the context I'm looking at it in)? I understand why it's not linearly ...
5
votes
2answers
299 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 ...
2
votes
2answers
506 views

If uniform cost search is used for bidirectional search, is it guaranteed the solution is optimal?

If uniform cost search is used for both the forward and backward search in bidirectional search, is it guaranteed the solution is optimal?
6
votes
1answer
149 views

Why does a negative reward for every step really encourage the agent to reach the goal as quickly as possible?

If we shift the rewards by any constant (which is a type of reward shaping), the optimal state-action value function (and so optimal policy) does not change. The proof of this fact can be found here. ...
0
votes
1answer
109 views

What is the optimal value function of the shifted version of the reward function?

Similarly to this question that I asked some time ago, what is the optimal value function of the shifted (by some constant $c$) version of some reward function? More precisely, let's assume that $r(s, ...
3
votes
2answers
71 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
25 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?
1
vote
0answers
30 views

How can I derive n-step off-policy temporal difference formula?

I was reading the book "Reinforcement Learning: An Introduction" by Sutton and Barto. In section 7.3, they write the formula for n-step off-policy TD as $$V(S_t) = V(S_{t-1}) + \alpha \rho_{...
0
votes
0answers
59 views

Can a computer make a proof by induction?

Can a computer solve the following problem, i.e. make a proof by induction? And why? Prove by induction that $$\sum_{k=1}^nk^3=\left(\frac{n(n+1)}{2}\right)^2, \, \, \, \forall n\in\mathbb N .$$ I'm ...
5
votes
2answers
141 views

Why are the Bellman operators contractions?

In these slides, it is written \begin{align} \left\|T^{\pi} V-T^{\pi} U\right\|_{\infty} & \leq \gamma\|V-U\|_{\infty} \tag{9} \label{9} \\ \|T V-T U\|_{\infty} & \leq \gamma\|V-U\|_{\infty} \...
6
votes
0answers
154 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', ...
1
vote
0answers
28 views

What is the proof that the variance of the gradient estimate in Actor-Critic is smaller than in REINFORCE?

The intuition provided when introducing actor-critic algorithms is that the variance of its gradient estimates is smaller than in REINFORCE as, e.g., discussed here. This intuition makes sense for the ...
1
vote
1answer
48 views

When to use AND and when to use Implies in first-order logic?

I am trying to learn the theory behind first-order logic (FOL) and do some practice runs of converting statements into the form of FOL. One issue I keep running into is hesitating on whether to use an ...
2
votes
1answer
214 views

Proof of Maximization Bias in Q-learning?

In the textbook "Reinforcement Learning: An Introduction" by Richard Sutton and Andrew Barto, the concept of Maximization Bias is introduced in section 6.7, and how Q-learning "over-estimates" action-...
2
votes
0answers
106 views

What is the proof that “reward-to-go” reduces variance of policy gradient?

I am following the OpenAI's spinning up tutorial Part 3: Intro to Policy Optimization. It is mentioned there that the reward-to-go reduces the variance of the policy gradient. While I understand the ...
1
vote
0answers
64 views

How do you prove that minimax algorithm outputs a subgame-perfect Nash equilibrium?

At every node, MAX would always move to maximise the minimum payoff while MIN choose to minimise the maximum payoff, hence there is nash equilibrium. By using backwards induction, at every node, MAX ...
5
votes
1answer
151 views

How do I convert an MDP with the reward function in the form $R(s,a,s')$ to and an MDP with a reward function in the form $R(s,a)$?

The AIMA book has an exercise about showing that an MDP with rewards of the form $r(s, a, s')$ can be converted to an MDP with rewards $r(s, a)$, and to an MDP with rewards $r(s)$ with equivalent ...
5
votes
1answer
152 views

Can deep learning be used to help mathematical research?

I am currently learning about deep learning and artificial intelligence and exploring his possibilities, and, as a mathematician at heart, I am inquisitive about how it can be used to solve problems ...
1
vote
0answers
89 views

Why is it hard to prove the convergence of the deep Q-learning algorithm?

Why is it hard to prove the convergence of the DQN algorithm? We know that the tabular Q-learning algorithm converges to the optimal Q-values, and with a linear approximator convergence is proved. ...
1
vote
1answer
51 views

Why is probability that at least one hypothesis out of $k$ being consistent with $m$ training examples $k(1- \epsilon)^m$?

My question is actually related to the addition of probabilities. I am reading on computational learning theory from Tom Mitchell's machine learning book. In chapter 7, when proving the upper bound ...
1
vote
1answer
103 views

Monte Carlo epsilon-greedy Policy Iteration: monotonic improvement for all cases or for the expected value?

I was going through university slides and this particular slide is trying to prove that in a Monte Carlo Policy Iteration algorithm using an epsilon-greedy policy, the state Values (V-Values) are ...
5
votes
2answers
412 views

What is the proof that policy evaluation converges to the optimal solution?

Although I know how the algorithm of iterative policy evaluation using dynamic programming works, I am having a hard time realizing how it actually converges. It appeals to intuition that, with each ...
2
votes
1answer
33 views

Equivalence between expected parameter increments in “Off-Policy Temporal-Difference Learning with Function Approximation”

I am having a hard time understanding the proof of theorem 1 presented in the "Off-Policy Temporal-Difference Learning with Function Approximation" paper. Let $\Delta \theta$ and $\Delta \...
2
votes
1answer
47 views

Is the derivative of the loss wrt a single scalar parameter proportional to the loss?

I am wondering about the correlation between the loss and the derivative of the loss wrt a single scalar parameter, with the same sample. That means: considering a machine learning model with ...
4
votes
1answer
138 views

Does Rice's theorem prove safe AI is undecidable?

According to Wikipedia In computability theory, Rice's theorem states that all non-trivial, semantic properties of programs are undecidable. A semantic property is one about the program's ...
2
votes
1answer
183 views

How to prove $\mathcal H$ with VC dimension $d$ shatter all subsets with size less than $d-1$?

I was wondering that if a certain hypothesis class $H$ has a VC dimension $d$ over domain $X$ how to prove that $H$ will shatter all subsets of $X$ with size less than $d$ i.e $H$ will shatter $A \...
1
vote
3answers
135 views

How to prove that gradient descent doesn't necessarily find the global optimum?

How can I prove that gradient descent doesn't necessarily find the global optimum? For example, consider the following function $$f(x_1, x_2, x_3, x_4) = (x_1 + 10x_2)^2 + 5x_2^3 + (x_2 + 2x_3)^4 + ...
2
votes
0answers
86 views

Is an oracle that answers only with a “yes” or “no” dangerous?

I was thinking about the risks of Oracle AI and it doesn't seem as safe to me as Bostrom et al. suggest. From my point of view, even an AGI that only answers questions could have a catastrophic impact....
3
votes
1answer
243 views

Is the summation of consistent heuristic functions also consistent?

Imagine that we have a set of heuristic functions $\{h_i\}_{i=1}^N$, where each $h_i$ is both admissible and consistent (monotonic). Is $\sum_{i=1}^N h_i$ still consistent or not? Is there any proof ...
1
vote
0answers
55 views

Does SARSA(0) converge to the optimal policy in expectation if the Robbins-Monro conditions are removed?

The conditions of convergence of SARSA(0) to the optimal policy are : The Robbins-Monro conditions above hold for $α_t$. Every state-action pair is visited infinitely often The policy is greedy with ...
1
vote
2answers
104 views

What are the conditions for the convergence of SARSA to the optimal value function?

Is it correct that for SARSA to converge to the optimal value function (and policy) The learning rate parameter $\alpha$ must satisfy the conditions: $$\sum \alpha_{n^k(s,a)} =\infty \quad \text{and}...
1
vote
1answer
112 views

Does TD(0) prediction require Robbins-Monro conditions to converge to the value function?

Does the learning rate parameter $\alpha$ require the Robbins-Monro conditions below for the TD(0) algorithm to converge to the true value function of a policy? $$\sum \alpha_t =\infty \quad \text{...
1
vote
1answer
425 views

Is there a simple proof of the convergence of TD(0)?

Does anybody know a simple proof of the convergence of the TD(0) value function prediction algorithm?
2
votes
1answer
104 views

Is there a mathematical theory behind why MLP can classify handwritten digits?

I'm trying to really understand how multi-layer perceptrons work. I want to prove mathematically that MLP's can classify handwritten digits. The only thing I really have is that each perceptron can ...
2
votes
1answer
109 views

Why does KL divergence not satisfy the triangle inequality?

$D_{KL}=\sum_i p(x_i)log(p(x_i)/q(x_i)$ Also can't you make it satisfy the triangle inequality by taking the absolute value of the information at every point?
2
votes
0answers
55 views

How can I show that the VC dimension of the set of all closed balls in $\mathbb{R}^n$ is at most $n+3$?

How can I show that the VC dimension of the set of all closed balls in $\mathbb{R}^n$ is at most $n+3$? For this problem, I only try the case $n=2$ for 1. When $n=2$, consider 4 points $A,B,C,D$ and ...
2
votes
0answers
226 views

A problem about the relation between 1-oracle and 2-oracle PAC model

This problem is about two-oracle variant of the PAC model. Assume that positive and negative examples are now drawn from two separate distributions $\mathcal{D}_{+}$ and $\mathcal{D}_{-} .$ For an ...
2
votes
0answers
50 views

How can we prove this inequality, related to the generalization error, without using the Rademacher complexity?

This is an inequality on page 36 of the book Foundations of Machine Learning, but the author only states it without proof. $$ \mathbb{P}\left[\left|R(h)-\widehat{R}_{S}(h)\right|>\epsilon\right] \...
2
votes
0answers
40 views

Convert a PAC-learning algorithm into another one which requires no knowledge of the parameter

This is part of the exercise 2.13 in the book Foundations of Machine Learning (page 28). You can refer to chapter 2 for the notations. Consider a family of concept classes $\left\{\mathcal{C}_{s}\...
3
votes
1answer
75 views

Why is the stationary distribution independent of the initial state in the proof of the policy gradient theorem?

I was going through the proof of the policy gradient theorem here: https://lilianweng.github.io/lil-log/2018/04/08/policy-gradient-algorithms.html#svpg In the section "Proof of Policy Gradient ...
3
votes
1answer
105 views

Understanding proof of lemma 1 (policy improvement bound) of the “Trust Region Policy Optimization” paper

In the Trust Region Policy Optimization paper, in Lemma 1 of Appendix A, I did not quite understand the transition from (21) from (20). In going from (20) to (21), $A^\pi(s_t, a_t)$ is substituted ...