# minimal_policy_gradient

**Repository Path**: dragon515/minimal_policy_gradient

## Basic Information

- **Project Name**: minimal_policy_gradient
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-07-16
- **Last Updated**: 2025-07-16

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

A minimal implementation of Policy Gradient in reinforcement learning. 

中文视频教程：【大白话强化学习之 Policy Gradient】 https://space.bilibili.com/90303434/lists/4898073?type=season

Our target is to keep the code easy enough to read and understand the theory behind gradient policy mathematical derivation:

![](policy_gradient_derivation.png)



# Training

call `train()` to start training. Press `S` to save a checkpoint, press `Q` to abort the training.

We are using CartPole-v0 as an example to train. The training should converge to max steps (500 steps) around **200 to 1000** steps.

![training](training.png)



# Evaluating

call `eval()` to start evaluate a checkpoint. Press `Q` to abort evaluating.



# Thanks

A great thanks to:

https://huggingface.co/learn/deep-rl-course/unit4/pg-theorem

https://github.com/Finspire13/pytorch-policy-gradient-example