What might be a step forward is a joint learning of the state representation with the search algorithm. Search algorithm explores the NN representation of the state for which you can get the cost.
Genie from DeepMind is a good demonstration where discrete state is being modeled. NN learns a very complex representation with collision detection and actions. Instead of decoding that state into pixels, search could probably be done directly on that state.
Of course, this architecture could be very different.
Active Inference is kinda "a joint learning of the state representation with the search algorithm"...or at least, related to the idea. I like framing it as a joint learning problem.
What might be a step forward is a joint learning of the state representation with the search algorithm. Search algorithm explores the NN representation of the state for which you can get the cost.
https://sites.google.com/view/genie-2024/
Genie from DeepMind is a good demonstration where discrete state is being modeled. NN learns a very complex representation with collision detection and actions. Instead of decoding that state into pixels, search could probably be done directly on that state.
Of course, this architecture could be very different.