Create custom policy function for a RL DQN.

Hi Community,

I am working on a project that requires me to have a little bit modification of the DQN policy. The learned function is still Q, but instead of taking the argmax Q(s,a), I have a few more conditions added (most likely some if statement as hard constraints). I am wondering if it is ever possible for me to make this change? If so, where should i work on?

Best regards,

Yiyang

Answer 1

[Anh Tran]

Currently there I do not see any workaround to modify DQN policy directly with buit-in rlDQNAgent. A possible workaround is to reimplement DQN agent with rlQValueRepresentation, introduced in MATLAB R2020a

You can refer to RL custom train loop example where we implement vanilla policy gradients with RL Toolbox.

For discrete action, I would recommend multi-output Q value representation Q(o) (better performance than Q(o,a)).

% create Q(o) critic, assumed you defined NeuralNet,ObservationInfo,ActionInfo
Critic = rlQValueRepresentation(NeuralNet,ObservationInfo,ActionInfo,'Observation',ObsLayerName);
% get state-action values of an observation RandomObservation
Q = getValue(Critic,RandomObservation)