After reading the early and the latter paper of NEAT, I still am not clear on the feature of NEAT for POMDP problem.
For instance, the game double-pole need velocity as a core property to solve. Without it , it is a POMDP problem.
As far as I know, if deep reinforcement learning(DRL) is taken to solve a POMDP, hiddent state must be taken as another input for network with RNN structure. And
the hidden state is the output from former-step of the network.
Butt... According to NEAT, it seems that only standard input like in standard MDP in needed for network with RNN structure (which is envolved by itself)---Is this a correct explanation?
If so, does it mean that NEAT can solve POMDP without building complex RNN structure, and without changing standard state input (on stand of standard MDP)?
After reading the early and the latter paper of NEAT, I still am not clear on the feature of NEAT for POMDP problem.
For instance, the game double-pole need velocity as a core property to solve. Without it , it is a POMDP problem. As far as I know, if deep reinforcement learning(DRL) is taken to solve a POMDP, hiddent state must be taken as another input for network with RNN structure. And the hidden state is the output from former-step of the network. Butt... According to NEAT, it seems that only standard input like in standard MDP in needed for network with RNN structure (which is envolved by itself)---Is this a correct explanation?
If so, does it mean that NEAT can solve POMDP without building complex RNN structure, and without changing standard state input (on stand of standard MDP)?