Next semester, I will teach a short course on Probability for university students who have not taken probability before, who know some basic mathematics, but who are not necessarily going to be studying mathematic.

The notes for this are here:

Skip to content
# Category: Uncategorized

## Probability: a short introduction

## Slides

## Neural Tangent Kernel

## Inventory Control

## Lai Robbins Lower-Bound

## Euler-Maruyama

## Stochastic Finite Arm Bandits

## Perturbation of Markov Matrices

## Cross Entropy Method

## A Short Discussion on Policy Gradients in Bandits

Next semester, I will teach a short course on Probability for university students who have not taken probability before, who know some basic mathematics, but who are not necessarily going to be studying mathematic.

The notes for this are here:

Here are slides from this year’s stochastic control course:

The Neural Tangent Kernel is a way of understanding the training performance of Neural Networks by relating them to Kernel methods. Here we overview the results of the paper [Jacot et al. here]. The paper considers a deep neural network with a fixed amount of data and a fixed depth. The weights applied to neurons are initially independent and normally distributed. We take a limit where the width of each layer tends to infinity.

We consider the problem where there is an amount of stock at time . You can perform the action to order units of stock. Further the demand at time is . We assume is independent over . The change in the amount of stock follows the dynamic:

We continue the earlier post on finite-arm stochastic multi-arm bandits. The results so far have suggest that, for independent identically distributed arms, the correct size of the regret is of the order . We now more formally prove this with the Lai and Robbins lower-bound .

Continue reading “Lai Robbins Lower-Bound”The Euler-Maruyama scheme is a method of approximating an stochastic differential equation. Here we investigate two forms of error the scheme: the weak error and the strong error. The aim is to later we will cover Multi-Level Monte Carlo (MLMC) and related topics.

Continue reading “Euler-Maruyama”We discuss a canonical multi-arm bandit setting. The stochastic multi-arm bandit with finite arms and bounded rewards. We let index the set of arms. We let be the set of arms. If you play the arm at time , you receive rewards which are independent and identically distributed in . However, the distribution between arms may change.

We let be the mean of arm . We want to find the machine with highest mean reward.

Continue reading “Stochastic Finite Arm Bandits”We consider the impact on the stationary distribution of perturbing a Markov chain’s transition matrix. This argument due to Seneta. We show if and are probability matrices of a irreducible finite state Markov chains with stationary distributions and then we can show that

**Theorem.**

where is the total variation between the rows of .

Continue reading “Perturbation of Markov Matrices”The Cross Entropy Method (CEM) is a generic optimization technique. It is a zero-th order method, i.e. you don’t gradients.^{1} So, for instance, it works well on combinatorial optimization problems, as well as reinforcement learning.

This is a quick note on policy gradients in bandit problems. There are a number of excellent papers coming out on the convergence of policy gradients [1 ,2, 3]. I wrote this sketch argument a few months ago. Lockdown and homeschooling started and I found I did not have time to pursue it. Then yesterday morning I came across the following paper Mei et al. [1], which uses a different Lyapunov argument to draw an essentially the same conclusion [first for bandits then for general RL]. Although most of the time I don’t talk about my own research on this blog, I wrote this short note here just in case this different Lyapunov argument is of interest.

Further, all these results on policy gradient are [thus far] for deterministic models. I wrote a short paper that deals with a stochastic regret bound for a policy gradient algorithm much closer to projected gradient descent, again in the bandit setting. I’ll summarize this, as it might be possible to use it to derive stochastic regret bounds for policy gradients more generally. At the end, I discuss some issues around getting some proofs to work in the stochastic case and emphasize the point that is a good choice of lyapunov function [where $latex $q^\star$ is the probability of *not* choosing the optimal action.]