|
|
--- |
|
|
tags: |
|
|
- reinforcement-learning |
|
|
- q-learning |
|
|
- gymnasium |
|
|
- cliffwalking |
|
|
library_name: gymnasium |
|
|
license: apache-2.0 |
|
|
--- |
|
|
|
|
|
# CliffWalking Q-Learning Agent |
|
|
|
|
|
This repository contains a Q-learning agent trained on the **CliffWalking-v0** environment from **Gymnasium**. The agent learns to navigate the cliff, avoiding falling into the cliff zone while reaching the goal with minimal penalties. The Q-learning algorithm is implemented with epsilon-greedy exploration and updates the Q-table based on state-action-reward transitions. |
|
|
|
|
|
## Files: |
|
|
- `train.py`: The main script that trains the Q-learning agent. |
|
|
- `cliffWalking_qtable.npy`: The saved Q-table after training. |
|
|
- `replay.mp4`: A video of the agent's performance after training. |
|
|
|
|
|
## Training Details: |
|
|
- **Environment**: `CliffWalking-v0` (Gymnasium) |
|
|
- **Episodes**: 30,000 |
|
|
- **Learning Rate (α)**: 0.2 |
|
|
- **Discount Factor (γ)**: 0.97 |
|
|
- **Epsilon (ε)**: 0.2 (exploration vs exploitation trade-off) |
|
|
|
|
|
The agent starts by exploring the environment randomly and gradually learns the optimal path to avoid falling off the cliff while reaching the goal. |
|
|
|
|
|
## How to Run: |
|
|
|
|
|
### 1. Install Dependencies: |
|
|
|
|
|
Make sure you have the required packages installed: |
|
|
```bash |
|
|
pip install gymnasium numpy imageio[ffmpeg] |
|
|
``` |
|
|
|
|
|
### 2. Training the Agent: |
|
|
To train the agent, run the script train.py: |
|
|
|
|
|
```bash |
|
|
python train.py |
|
|
``` |
