_index.org

Policy Gradient

Last edited: August 8, 2025

Two steps:

  1. obtaining a function for the gradient of policy against some parameters \(\theta\)
  2. making them more based than they are right now by optimization

Thoughout all of this, \(U(\theta)\) is \(U(\pi_{\theta})\).

Obtaining a policy gradient

Finite-Difference Gradient Estimation

We want some expression for:

\begin{equation} \nabla U(\theta) = \qty[\pdv{U}{\theta_{1}} (\theta), \dots, \pdv{U}{\theta_{n}}] \end{equation}

we can estimate that with the finite-difference “epsilon trick”:

\begin{equation} \nabla U(\theta) = \qty[ \frac{U(\theta + \delta e^{1}) - U(\theta)}{\delta} , \dots, \frac{U(\theta + \delta e^{n}) - U(\theta)}{\delta} ] \end{equation}

policy iteration

Last edited: August 8, 2025

policy iteration will allow us to get an optimal policy.

  1. start with some initial policy \(\pi\) (this scheme converges to an optimal policy regardless of where you start)
  2. solve for \(U^{\pi}\)
  3. create a new policy \(\pi’\) by creating a value-function policy on \(U^{\pi}\)
  4. repeat 2-3

Since there are a finite policies, this will eventually converge.

At each point, the utility of the policy increases.

At each step, the utility of the resulting policy will necessarily be larger or equal to than the previous one as we are greedily choosing “better” (or equivalent) actions as measured by the utility of the previous policy.

Policy Optimization

Last edited: August 8, 2025

Policy Optimization deals with algorithms that, unlike value iteration/policy iteration/online planning which uses a surrogate (like value function or some future discounted reward) to calculate a policy, directly optimizes against policy parameters \(\theta\) for a policy \(\pi_{\theta}\).

Polio

Last edited: August 8, 2025

polynomial

Last edited: August 8, 2025

A polynomial is a polynomial

constituents

  • a function \(p: \mathbb{F} \to \mathbb{F}\)
  • coefficient \(a_0, \dots, a_{m} \in \mathbb{F}\)

requirements

A polynomial is defined by:

\begin{equation} p(z)=a_0+a_1z+a_2z^{2}+\dots +a_{m}z^{m} \end{equation}

for all \(z \in \mathbb{F}\)

additional information

degree of a polynomial \(\deg p\)

A polynomial’s degree is the value of the highest non-zero exponent. That is, for a polynomial:

\begin{equation} p(z) = a_0+a_1z+\dots +a_{m}z^{m} \end{equation}

with \(a_{m} \neq 0\), the degree of it is \(m\). We write \(\deg p = m\).