The learning algorithm used is vanilla Deep Q Learning as described in original paper. As an input the vector of state is used instead of an image so convolutional neural nework is replaced with deep neural network. The deep neural network has following layers:
- Fully connected layer - input: 37 (state size) output: 128
- Fully connected layer - input: 128 output 64
- Fully connected layer - input: 64 output: (action size)
Parameters used in DQN algorithm:
- Maximum steps per episode: 1000
- Starting epsilion: 1.0
- Ending epsilion: 0.01
- Epsilion decay rate: 0.999
Episode 100 Average Score: 1.66
Episode 200 Average Score: 8.17
Episode 300 Average Score: 10.81
Episode 364 Average Score: 13.00
Environment solved in 264 episodes! Average Score: 13.00
- Extensive hyperparameter optimization
- Double Deep Q Networks
- Prioritized Experience Replay
- Learning from pixels