https://github.com/codeamt/maze_rl

a multi-agent cooperative maze game using gym.
https://github.com/codeamt/maze_rl

Last synced: about 2 months ago
JSON representation

a multi-agent cooperative maze game using gym.

• Host: GitHub
• URL: https://github.com/codeamt/maze_rl
• Owner: codeamt
• Created: 2021-01-09T00:58:33.000Z (almost 4 years ago)
• Default Branch: master
• Last Pushed: 2021-01-09T06:04:21.000Z (almost 4 years ago)
• Last Synced: 2023-10-20T04:52:17.603Z (about 1 year ago)
• Language: Python
• Size: 149 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# maze_rl



A multi-agent cooperative maze game using Open AI's gym.





  



## About

This reinforcement learning (RL) game uses Proximal Policy Optimization (PPO2) to train two agents to cooperate in figuring there way out of a n*m (5x5) grid world (the "Maze"). 

**Rules:**


- Agents must stay within bounds  of the nxm Maze 

- Agents cannot occupy the same position simultaneously 

**Entities:**


- 0: Available Space

- 1: Agent 1 

- 2: Agent 2 

- 3: A Trap (Instant Loss)

- 4: A Teleportation Portal (Moves partner 3 steps North)

- 5: An Exit (A win) 

**Action Space (Discrete):** 


- 0: Up

- 1: Right

- 2: Down 

- 3: Left 

**Observation Space:**

- (np.array) -> next player and maze world, flattened, concatenated on horizontal axis.

**State:**

- P: In Play 

- L: Agents Lost 

- W: Agents Won 

**Reward Function/Incentive Mechanism:**

- Reward Range: (-200, 200) 

- Incentives: +1 for Exploring 

- Penalties: -2 for moving to a previopusly visited cell

## Getting started

Clone the repo:

```

https://github.com/codeamt/maze_rl.git

```

Install depemndencies:

```

cd maze_rl && pip install -r requirements.txt

```

## Running Episodes (Local)

Change into src directory and run script:

```

cd code && python main.py

```

### Args: 

- **--epochs** 


(type: int, default: 500000)


Number of epochs

- **--lr** 


(type: float, default: 0.001)


Learning rate for training policy

- **--gamma** 


(type: float, default: 0.000001)


Discount Factor 

- **--lam** 


(type; float, default: 0)


Lambda/GAE Factor

- **--world** (type: List[List[int],
 

default:
 

[[1, 0, 2, 0, 0],


[0, 0, 0, 0, 0],
 

[4, 0, 0, 0, 0],


[0, 0, 3, 3, 0],


[0, 0, 3, 5, 0]]

- **--inference** (type: bool, default: True)

Whether or not to run a quick inference test after training to test performance. 

After some warnings from Tensorflow, you will see the updated maze after each steps on the output.

Check the render folder for a training report.

## Docker Build/Run Instructions

Build the image:

```

docker build -t  .

```

Run the container: 

```

docker run -d -it --name=  python main.py --epochs=500000 

```