https://github.com/keon/deep-q-learning
Minimal Deep Q Learning (DQN & DDQN) implementations in Keras
https://github.com/keon/deep-q-learning
ddqn deep-learning deep-q-network deep-reinforcement-learning dqn reinforcement-learning
Last synced: 30 days ago
JSON representation
Minimal Deep Q Learning (DQN & DDQN) implementations in Keras
- Host: GitHub
- URL: https://github.com/keon/deep-q-learning
- Owner: keon
- License: mit
- Archived: true
- Created: 2017-02-06T02:57:42.000Z (over 9 years ago)
- Default Branch: master
- Last Pushed: 2020-12-04T22:57:25.000Z (over 5 years ago)
- Last Synced: 2025-01-18T22:38:52.597Z (over 1 year ago)
- Topics: ddqn, deep-learning, deep-q-network, deep-reinforcement-learning, dqn, reinforcement-learning
- Language: Python
- Homepage: https://keon.io/deep-q-learning
- Size: 1.77 MB
- Stars: 1,293
- Watchers: 64
- Forks: 453
- Open Issues: 18
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# deep-q-learning
Introduction to Making a Simple Game AI with Deep Reinforcement Learning
Minimal and Simple Deep Q Learning Implemenation in Keras and Gym. Under 100 lines of code!
The explanation for the `dqn.py` code is covered in the blog article
[https://keon.kim/writing/deep-q-learning/](https://keon.kim/writing/deep-q-learning/)
I made minor tweaks to this repository such as `load` and `save` functions for convenience.
I also made the `memory` a deque instead of just a list.
This is in order to limit the maximum number of elements in the memory.
The training might be unstable for `dqn.py`. This problem is mitigated in `ddqn.py`.
I'll cover `ddqn` in the next article.
## Changelog
### 2026-05
- Migrated from `gym` to `gymnasium` and updated for the modern reset/step API.
- Updated for Keras 3 (`Adam(learning_rate=)`, `Input` layer, `keras.losses.Huber`).
- `ddqn.py` now implements actual Double DQN (online net selects, target net evaluates).
- Removed CartPole-specific reward shaping from `ddqn.py` so it's env-agnostic.
- `save()` / `load()` now persist `epsilon` alongside the weights.
### 2018-08
- Split batched-update training into its own file (`dqn_batch.py`).
### 2017-06
- Added Double DQN with Huber loss in `ddqn.py`.
- Added `requirements.txt`.
### 2017-05
- Migrated to the Keras 2 API.
- Hyperparameter tuning that finally got `dqn.py` learning reliably on CartPole.
### 2017-02
- Initial release: DQN agent on CartPole with experience replay (`deque`-bounded memory) and `save` / `load` helpers.