The implemented double Q-learning algorithm utilizes two neural networks – primary and target
neural network. The primary network is trained using back-propagation, whereas the target network
is not. Instead, the target network is updated every x number of steps by setting its parameters to the
model parameters of the primary network that are learned through back-propagation. The primary
and target neural nets have the same architectures – they are both deep fully connected neural nets
with 64 neurons in the first layer, 20 neurons in the second layer and 3 neurons in the last layer – the
3 neurons in the last layer represent the three possible actions. The input dimension on the neural
nets is a 1-d array of six numbers and the non-linear activation function is relu. The last layer uses a
linear action function.
The envirnoment the algorithm is tested in is Acrobot-v1 taken from the OpenAI Gym package.