Double DQN

Used a 2-layer Fully Connected network with H1=100, H2=60 and ReLU

He Initialization of weights
Adam Optimizer. Initial learning rate = 0.001 Learning rate reduced using gamma of 0.50 every 350 episodes

Gamma = 0.99
Eps   = 1.00
Eps Decay = 0.98
Eps Decay every new episode. (not each step)

Uniform sampling from a replay buffer with 150,000 memories. No learning for first 1500 steps. (to fill replay buffer).
I tried to implement prioritized experience replay, but couldn't get it to work. (yet)

States normalized based on emperical mean and std of historical observation.

No reward or gradient clipping.

Target Network updated every 600 steps

Smooth L1 Loss (Huber Loss) rather than MSE.

Implemented in pyTorch and python 3.5. Trained on GTX-1070