FindGem

FindGem_v1 Use DQN net

FindGem_v1 is a game that the robot walks in a two-dimensional maze. There are fire pits, stone pillars, and diamonds in the maze. If the robot falls into the fire pit, the game is over. If you find diamonds, you can get rewards and the game is over! Design the best strategy to let the robot find the diamond as soon as possible and get rewards.

Every time the environment is reset, it will only change the position of the robot and not the positions of gems, fire pits, and stone pillars

---

Result Show

• video

• gif
  

gym Env set up

FindGem's Env setting Reference from this link Based on the original author, made appropriate bug fixes and robot move position modifications

pip install -r requirements.txt

1. First move Custom env code move to your site-package gym path /anaconda3/Lib/site-packages/gym/envs

mkdir /anaconda3/Lib/site-packages/gym/envs/user
cp ./env/grid_findGem_v1.py /anaconda3/Lib/site-packages/gym/envs/user
cp ./env/__init__.py /anaconda3/Lib/site-packages/gym/envs/user

2. Registration environment in gym open .../anaconda3/Lib/site-packages/gym/envs/__init__.py add the following code

 register(
 id='GridWorld-v1',
 entry_point='gym.envs.user:GridEnv1',
 max_episode_steps=200,
 reward_threshold=100,
 )

Train Model

DQN model reference from pytorch official tutorials train process result save reference from tensorflow org tutorials and Modify the reward mechanism, increase with the number of steps, reduce rewards, and update status acquisition

  python customEnvFindGem.py --mode train --BATCH_SIZE 256  --num_episodes 50000 

model file

Test model

  python customEnvFindGem.py --mode test 

result file will save at ./result/result_polyDL.mp4 result_mp4