first version of training loop for dyna

[Misterhoonster]

Could you take a look at my starter training loop logic?

[wcarvalho]

Have a single Network with

1. setup
2. initialize_carry
3. call for when its used with read data
4. apply_model

[wcarvalho]

call: takes in observation, applies rnn, returns state + predictions(q_vals)
apply_model: takes in state and action, applies model (for us environment, for somebody else a neural network), returns next_state + predictions(q_vals)

[wcarvalho]

call(self, observation) --> [AgentState, Predictions]

apply_model(self, state, action) --> [AgentState, Predictions]

[wcarvalho]

Think of this as the Agent

[wcarvalho]

call: takes in observation, applies rnn, returns state + predictions(q_vals)
apply_model: takes in state and action, applies model (for us environment, for somebody else a neural network), returns next_state + predictions(q_vals)

[wcarvalho]

call(self, observation) --> [AgentState, Predictions]

apply_model(self, state, action) --> [AgentState, Predictions]

[wcarvalho]

Think of this as the Agent

[Misterhoonster]

Hey Wilka, just updated my dyna file with your suggested changes!

[Misterhoonster]

Pushed the new version with the Loss class! Can you take a look, please?