Mathematics for AI and Machine Learning

Foundations for modern AI and machine learning

Policy Gradient

Chapter 15 Reinforcement learning

Chapter 15: Bellman Equations and Operators — Policy Gradient Theorem & Stochastic Approx.

Animated preview (GIF)
Policy Gradient — animated GIF preview
High-resolution PNG
Policy Gradient — high-resolution mind-map icon

From the book

Chapter 15: Bellman Equations and Operators. In the chapter mind map this icon labels Policy Gradient Theorem & Stochastic Approx.. The discussion below is excerpted and lightly edited from § Policy Gradient Theorem & Stochastic Approx. in Mathematics for AI and Machine Learning.

The gradient of the expected return with respect to policy parameters is:

where $\mu^{\pi_\theta}$ is the stationary state distribution under policy $\pi_\theta$.

For the infinite-horizon case with discounting, this becomes:

where $d_{\pi_\theta}(s) = (1-\gamma) \sum_{t=0}^{\infty} \gamma^t \mu_t(s)$ is the discounted state visitation distribution.

What this drawing shows

What you see. Purple policy simplex $\pi_\theta$ and red $\nabla_\theta J$ arrow fixed; a blue policy parameter moves uphill along the gradient update.

In the mind map. Chapter 15 — Policy Gradient Theorem & Stochastic Approx.. See From the book above for definitions, figures, and worked examples.

Where to read next

Open Chapter 15 companion →

Read the full definitions, figures, and worked examples in Chapter 15: Bellman Equations and Operators — see the mind-map node Policy Gradient Theorem & Stochastic Approx..