https://github.com/amanfoundongithub/double-dqn
Implementation of Double DQN network for stable training
https://github.com/amanfoundongithub/double-dqn
double-dqn dqn
Last synced: about 1 month ago
JSON representation
Implementation of Double DQN network for stable training
- Host: GitHub
- URL: https://github.com/amanfoundongithub/double-dqn
- Owner: amanfoundongithub
- Created: 2025-05-20T02:24:13.000Z (5 months ago)
- Default Branch: master
- Last Pushed: 2025-05-22T13:53:44.000Z (5 months ago)
- Last Synced: 2025-06-24T04:12:03.486Z (4 months ago)
- Topics: double-dqn, dqn
- Language: Python
- Homepage:
- Size: 203 KB
- Stars: 2
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Double DQN
Implemented the Double DQN ([David Silver et. al](https://arxiv.org/pdf/1509.06461)), a variant of DQN that uses two networks to reduce the over-estimation error from DQN by using a target DQN to keep the updates within the limits of the original DQN.
## Environment
Implementation was tested on `LunarLander-v2` environment of the OpenAI Gym.
## Result of Training
The epsilon was also decreased linearly from 1 to 0.1 in 50,000 steps.
Double DQN shows a strong learning curve when we train it. We trained it on 1,00,000 steps i.e. almost 200 episodes (each episode contains at max 500 steps).
- Initially, it got very negative rewards due to large epsilon. However, the negative values gradually decreased.
- After 35,000 frames it started to earn positive overall reward for the first time.
- In 50,000 frames, it almost earned >100 reward.
- There were regular cases with negative rewards, which indicates that the agent was learning.
- The plots:
## Test
We tested the agent and it scored 102.0 in an framework.
## Result
Double DQN is way more faster and stable than DQN. In less than 200 episodes, we got almost 100 reward which is way more faster than DQN.