# Stanford AA228/CS238 Decision Making Under Uncertainty I Policy Gradient Estimation and Optimization | Summary and Q&A

7.6K views
November 7, 2023
by
Stanford Online
Stanford AA228/CS238 Decision Making Under Uncertainty I Policy Gradient Estimation and Optimization

## TL;DR

This content provides an overview of policy optimization and introduces the likelihood ratio policy gradient, which is an unbiased estimator for the gradient of a policy's expected utility.

## Install to Summarize YouTube Videos and Get Transcripts

### Q: What is policy optimization and why is it important?

Policy optimization involves making decisions under uncertainty and maximizing the utility of a set of actions in an environment. It is important because it allows us to find the best strategy (policy) for an agent to take actions that optimize its performance in a given context.

### Q: How does the likelihood ratio policy gradient relate to policy search?

The likelihood ratio policy gradient is a method used in policy search to estimate the gradient of a policy's expected utility. It provides a way to optimize the policy by quantifying how actions should be adjusted in various states to maximize the expected utility.

### Q: What is the advantage of working with stochastic policies instead of deterministic policies?

Stochastic policies consider a distribution of actions given a state, which allows for exploration and adaptation to uncertain environments. Deterministic policies, on the other hand, always output the same action for a given state. Stochastic policies are more flexible and can better handle uncertainty.

### Q: How does the reward-to-go approach help to reduce variance in the likelihood ratio policy gradient estimation?

The reward-to-go approach focuses on the rewards obtained when taking an action and disregards past rewards. By considering only the future rewards, it reduces the variance in the estimation of the gradient by eliminating the influence of past actions on current rewards.

## Summary & Key Takeaways

• Policy optimization involves making decisions under uncertainty and optimizing a series of actions in an environment.

• The policy is a strategy that guides the agent's actions in a given state.

• The likelihood ratio policy gradient is an unbiased estimator for the gradient of the policy's expected utility and helps to improve the policy optimization process.