Deep Q Networks

Generic implementation of Deep Q Networks using Keras for the purpose of solving RL control tasks having discrete action space.

Requirements :

numpy
matplotlib
keras
imageio
gym (OpenAI Gym)

Help :

$ python main_driver -h

Environments :

List of OpenAI Gym Environments which are solved by agents in this repo:

• CartPole-v0

  Targeted Reward (over 100 consecutive itrs) : >= 195
  Reward Achieved (over 100 consecutive itrs) : 200.0
  Episodes before solve : 265
  

  Following are the commands used to train & test the model(includes the necessary hyperparameters):

  • Command to train : python3 main_driver,py train CartPole-v0 --itr-count 2000 -b 64 -lr 0.0001 --gamma 0.99 --render False --avg-expected-reward -110.0

  • Command to test : python3 main_driver,py test CartPole-v0 --render True --avg-expected-reward 195 --test-total-iterations 100 --model-path "best_save/CartPole-v0_local_model_1534607382.5808585.h5"

  Result Log Files:

  • Training : result_train_CartPole-v0_1534607382.5808585.log
  • Testing : result_test_CartPole-v0_1534607382.5808585.log

  Detailed Log Files:

  • Training : train_CartPole-v0_1534607382.5808585.log
  • Testing : test_CartPole-v0_1534607382.5808585.log

  Reward Plot:
  

  Agent Playing the Env:
  

---

• MountainCar-v0

  Targeted Reward (over 100 consecutive itrs) : >= -110
  Reward Achieved (over 100 consecutive itrs) : -108.85
  Episodes before solve : 643
  

  Following are the commands used to train & test the model(includes the necessary hyperparameters):

  • Command to train : python3 main_driver,py train MountainCar-v0 --itr-count 2000 -b 64 -lr 0.0001 --gamma 0.99 --render False --avg-expected-reward -110.0

  • Command to test : python3 main_driver,py test MountainCar-v0 --render True --avg-expected-reward -110 --test-total-iterations 100 --model-path "best_save/MountainCar-v0_local_model_1534667457.5695965.h5"

  Result Log Files:

  • Training : result_train_MountainCar-v0_1534667457.5695965.log
  • Testing : result_test_MountainCar-v0_1534667457.5695965.log

  Detailed Log Files:

  • Training : train_MountainCar-v0_1534667457.5695965.log
  • Testing : test_MountainCar-v0_1534667457.5695965.log

  Reward Plot:
  

  Agent Playing the Env: