nitish-kalan / mountaincar-v0-deep-q-learning-dqn-keras

Solving MountainCar-v0 environment in Keras with Deep Q Learning an Deep Reinforcement Learning algorithm

MountainCar-v0 with Deep Q-Learning in Keras

MountainCar-v0 is an environment presented by OpenAI Gym. In this repository we have implemeted Deep Q Learning algorithm [1] in Keras for building an agent to solve MountainCar-v0 environment.

Commands to run

• To train the model

  python train_model.py
  

• To test the model

  python test_model.py 'path_of_saved_model_weights' (without quotes)
  

• To test agent with our trained weights

  python test_model.py saved_model/-134.0_agent_.h5
  

The greater the value of the episodic reward achieved, the better is the model.

Results

• Output of agent taking random actions

  

• Output of our agent at Episode: 550 with score -134.0

  

References

[1] Playing Atari with Deep Reinforcement Learning Authors: Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, Martin Riedmiller
Link: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf

mountaincar-v0-deep-q-learning-dqn-keras's Issues

Why don't you use two networks for stabilization?

Hi,

I thought DQN would be tricky regarding convergence and that you more or less need the fixed q-targets and experience replay as workaround. However, looking at your code I see a replay buffer but you use the same network for action choice online and adjusting the one-step lookahead update of your q-values. Is there any reason for that?