Gary Marcus

also you are ignoring all the work the monte carlo tree search and related infrastructure contributed, apparently attributing the entirely solution to deep learning, which would be inaccurate

Not at all, I'm quite the fan of MCTS, but it isn't a variable. The convolutions were a key bit of prior knowledge, but also weren't variables. You claimed that symbolic manipulation would be required to move past perception, but I fail to see any symbolic representations here.

mcts is a tree-governed system

MCTS is a search procedure, whereas you seem to be mistaking it for tree-structured representations. That they both have "tree" in the name seems besides the point.

the procedure operates over variables, with lots of variable binding taking place. that’s what matters. (and as it happens at least in the original Nature paper parts of the tree itself were encoded, if i recall correctly.)

That is a very liberal definition of variable binding. That would encompass all planning procedures (e.g. s_{t+n} is bound to the state predicted in n steps). If wading into model-based RL is all you wanted, then every DRL researcher is already on board. Mission accomplished?

i don’t see DQN itself that way.

You are in a node (variable) and there is a number of alternatives for the next move. You switch to a new node (variable changed) by selecting an action. Thus you manipulated a variable. Does deep learning make you that blind?