Gary Marcus

Many people have tried to defend pure DRL w Nature Machine Intelligence article that actually is a hybrid model; the below youtube pointer IS to a pure convnet - but read fine print: “network works decently well for any position less than 6 moves away from solved” #symbolphobiahttps://twitter.com/AlexRoseGames/status/1186571935611850752 …

To solve Rubik's cube in generality, you need to approach it as a reinforcement learning problem. You must explore the search space somehow, e.g., MCTS, Thompson sampling, etc. Is all search in discrete spaces "symbolic"? Not in an interesting way

I doubt this is how most people learn it. One reads books (or watches youtube) and learns the heuristics (mostly if-then rules) by memorization.

As to learning the solve without instruction, I would expect that people use a mixture of forward modelling and RL. It can't just be state space exploration.

And the forward model is learned somehow. Perhaps through exploring the state space? It's interesting that the model can be factored into the rotations of the cube and the locations of the colored patches. AFAIK that's beyond the state of the art for RL

Rubik's cube is a great challenge for creating an agent that can start by learning a forward model and then compile it (through practice == RL) into pattern->action rules. Do we have any learning method that can do all of this in a general way?