使用神经网络将旧系统更新为Q学习

Question

Recently I've been reading a lot about Q-learning with Neural Networks and thought about to update an existing old optimization system in a power plant boiler composed of a simple feed-forward neural network approximating an output from many sensory inputs. The output then is linked to a linear model-based controller that somehow output again an optimal action so the whole model can converge to a desired goal.

Identifying linear models is a consuming task. I thought about refurbishing the whole thing to model- free Q-learning with a Neural Network approximation of the Q-function. I drew a diagram to ask you if I'm on the right track or not.

My question: if you think I understood well the concept, should my training set be composed of State Features vectors from one side and Q_target - Q_current (here I'm assuming there's an increasing reward) in order to force the whole model towards the target or am I missing something?

Note: The diagram shows a comparison between the old system in the upper part and my proposed change on the lower part.

EDIT: Does a State Neural Network guarantee Experience Replay?

Answer 1

You might just use all the Q value of all the actions in the current state as the output layer in your network. A poorly drawn diagram is here

You can therefore take advatange of NN's ability to output multiple Q value at a time. Then, just back prop using loss derived by Q(s, a) <- Q(s, a) + alpha * (reward + discount * max(Q(s', a')) - Q(s, a) , where max(Q(s', a')) can be easily computed from the output layer.

Please let me know if you have further questions.