hsuRush / DeepDoodleJump

Playing DoodeJump with DQN and DRQN.

Using Reinforcement Learning to Learn How To Play Doodle Jump

Overview

This project follows the description of the Deep Q Learning algorithm described in Playing Atari with Deep Reinforcement Learning [1] and shows that this learning algorithm can be further generalized to the doodle jump.

Furthermore, I implement Deep Recurrent Q Learning [2] and outperform the original DQN a lot.

Installation Dependencies:

• Python 3
• TensorFlow
• pygame
• OpenCV-Python

Experiment

• Since deep Q-network is trained on the raw pixel values observed from the game screen at each time step. I remove the background appeared in the original game to make it converge faster.
• Also, To make the network train faster, I close off the red and blue platforms with only green platforms remain.

Preprocessing

Terminated Problem (Non-stop)

• Solution: The game will be terminated if there's no veritcal camera movement in over 5 seconds.

DQN Architecture

• I Stack 4 input images as a input tensor

DRQN Architecture

• I Stack only 1 input images as a input tensor
• replace a Dense layer as a LSTM layer.

args

• --actions 2: actions contains [LEFT, RIGHT] 3: actions contains [LEFT, RIGHT, DO_NOTHING]

• -- exp_dir the folder that contains model weights.

args only for TESTING

• --iterations
  testing in iterations and it count by death.
• --fps for visualizing the testing process, it will slow down the inferencing so not recommend for recoding benchmark.

Training DQN

python train_with_dqn.py --actions 2 --exp_dir exp_dqn_1

Training DRQN

python train_with_drqn.py --actions 2 --exp_dir exp_drqn_1

Testing DQN

• --fps 60 if you want to visualize the process, else remove --fps 60.

python test_with_dqn.py --actions 2 --exp_dir exp_drqn --fps 60

Testing DRQN

• --fps 60 if you want to visualize the process, else remove --fps 60.

python test_with_drqn.py --actions 2 --exp_dir exp_drqn --fps 60

Result

• DRQN: trained 1285,0000 iterations
• DQN : trained 1605,0000 iterations
• test until 100 death.

_	DQN	DRQN
Survived time per death(s)	0.77	10.3
top score per death	17	430
avg score in 100 death	3	96

---

Updated (6/18)

• DRQN: trained 1960,0000 iterations
• DQN : trained 1605,0000 iterations
• test until 100 death.

_	DQN	DRQN
Survived time per death(s)	0.77	22.79
top score per death	17	1464
avg score in 100 death	3	256

Conclusion

• The hidden reward platform (①,②) to the real reward(③) can be seen as a time sequence reward, the RNN is great at solving time sequence issues.

References

[1] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing Atari with Deep Reinforcement Learning. NIPS, Deep Learning workshop

[2] Matthew Hausknecht and Peter Stone. Deep Recurrent Q-Learning for Partially Observable MDPs. AAAI 2015

[3] https://github.com/f-prime/DoodleJump

[4] https://github.com/yenchenlin/DeepLearningFlappyBird