ManySpace
Türkçe Sign in Sign up

AI & MLRecommendedAIML-RL-016

Policy Gradients: REINFORCE

Gradient ascent on expected return using the log-derivative trick; baselines reduce variance.

Difficulty
3 / 5
Importance
4 / 5
Study time
1–2 h
Kind of work
Mixed

Why it matters

The learner can derive the REINFORCE update for a simple policy and explain how subtracting a baseline reduces variance without adding bias.

The route to this topic

Route AIML-RL-016
What Is a Model? Rational Agents and Task Environments Features and Targets Agent and Environment Linear Regression I: Prediction and Residuals Return and Discounting Policies Loss and Fitting: Squared Error Markov Decision Processes Objective Functions and Empirical Risk Policy Optimization Gradient Descent in Machine Learning Stochastic and Mini-batch Optimization Policy Gradients: REINFORCE