# UC-Berkeley-CS-294-Deep-RL

**Repository Path**: colin-cheng/UC-Berkeley-CS-294-Deep-RL

## Basic Information

- **Project Name**: UC-Berkeley-CS-294-Deep-RL
- **Description**: UC-Berkeley-CS-294-Deep-RL
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-09-01
- **Last Updated**: 2022-09-02

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# CS 294 at UC Berkeley: Deep Reinforcement Learning

This repo includes my solutions to the assigments of the UC Berkeley Deep Reinforcement Learning course offered in Fall 2018, taught by Sergey Levine. 

Course URL: http://rail.eecs.berkeley.edu/deeprlcourse/ 

Course GitHub: https://github.com/berkeleydeeprlcourse/homework

I would like to thank the instructor and the TAs for making this wonderful course.

## Code structure
- `./hw1/`: contains code for the implementation of **imitation learning**, including direct **behavior cloning** and the **DAgger** algorithm. See my report for this homework [here](./hw1/README.md). 
- `./hw2/`: contains code for the implementation of **policy gradient** and its variants, including **variance reduction methods**. See my report for this homework [here](./hw2/README.md). Below you can see some cool videos showing the comparison between the learning agent before and after training using the policy gradient.

|  |Before training  |	After training 	
|--|--|--|	
| CartPole | <img src="./videos/CartPole-wo-learning.gif" height="300px" width="300px" > | <img src="./videos/CartPole-w-learning.gif" height="300px" width="300px" > |
| LunarLander | <img src="./videos/LunarLander-wo-learning.gif" height="300px" width="300px" > | <img src="./videos/LunarLander-w-learning.gif" height="300px" width="300px" > |
| InvertedPendulum | <img src="./videos/InvertedPendulum-wo-learning.gif" height="300px" width="300px" > | <img src="./videos/InvertedPendulum-w-learning.gif" height="300px" width="300px" > |
| HalfCheetah | <img src="./videos/HalfCheetah-wo-learning.gif" height="300px" width="300px" > | <img src="./videos/HalfCheetah-w-learning.gif" height="300px" width="300px" > |

- `./hw3/`: contains code for the implementation of **Deep Q-Learning** and **Actor-Critic**. See my report for this homework [here](./hw3/README.md). Below you can see some cool videos showing the comparison between the learning agent before and after training using Deep Q-Learning.

|  |Before training  |	After training 	
|--|--|--|	
| LunarLander | <img src="./videos/LunarLander-wo-learning-ddqn.gif" height="300px" width="300px" > | <img src="./videos/LunarLander-w-learning-ddqn.gif" height="300px" width="300px" > |

`Pong Atari (green pedal is controlled by our learning agent) `

| After 0 timesteps | 	After 500K timesteps | After 1M timesteps | After 1.5M timesteps | After 2M timesteps
|--|--|--|--|--|
| <img src="./videos/Pong_0K.gif" height="200px" width="200px" > |<img src="./videos/Pong_500K.gif" height="200px" width="200px" > |<img src="./videos/Pong_1000K.gif" height="200px" width="200px"> |<img src="./videos/Pong_1500K.gif" height="200px" width="200px"> |<img src="./videos/Pong_2000K.gif" height="200px" width="200px" >|

- `./hw4/`: contains code for the implementation of an algorithm for the **model-based reinforcement learning (MBRL)**. See my report for this homework [here](./hw4/README.md).