Gary Marcus

Both can be & are true. Really telling to me is that DRL on its own worked for Atari games but not Go— and that DM’s spin on Go really downplayed the hybrid aspect that was essential to its success. (It was also apparently necessary to build in the rules for Go, unlike Atari.)

So we build hybrid systems but are also too reliant on DRL? How can both of those be true?

It’s question of emphasis, in part, but if I were running your ship I would spend more time exploring principled ways of building hybrids, and more kinds of of hybrids, and more time on on open-ended problems.

Other than Neural Turing Machines, AlphaGo, GraphNets, GQN, SPIRAL, etc? I'm sure you'd run things differently, but this is a far cry from the DRL centric narrative of the Wired article.

Oh, so now AlphaGo is a hybrid? :) but yes I like a lot of that work and have advocated for some of it over time. I totally agree that DRL is not the only emphasis at DM; it’s just the largest (from what I can tell) and my least favorite and most visible, wrapped in one.

It's what you'd consider a hybrid. To RL folks the jump from DQN to MCTS doesn't change fields, so "hybrid" sounds weird.

Its not weird, it’s what (in conjunction) with RL makes it work. You have drunk the KoolAid if you ignore this.

MCTS is literally in RL 101 textbooks. The tree structure of MDPs is inherent in their temporal structure and core to RL. I'm sure it reminds you of Chomsky, parse trees, grammars, and that's fine, but not invoking them isn't "drinking the KoolAid", it's understanding the field.