[civilian]

I love the idea of cognitive AI. I've dabbled with OpenCog https://en.wikipedia.org/wiki/OpenCog , and in my youth I read "Artificial General Intelligence (2005)" by Ben Goertzel and I was really convinced.

But it seems like... cognitive AI hasn't paid off? Big & deep Neural Nets are the type of ML/AI that are achieving milestones in learning, gameplay and tasks.

If someone has a strong case to make for cognitive programming, I'd love to hear it. But right now it seems like it's a heuristics-based system that's destined to lose to NNs.

(And before the nitpickers arrive, I totally grant that using heuristics for toy robots makes sense, and is a good way to expose people to programming.)

[mtlmtlmtlmtl]

FWIW, classical algorithms(with a small neural net for eval) is still the strongest approach for chess, consistently beating out more heavily NN based approaches in TCEC. And I'm not an expert, but I'm pretty sure the strongest AIs for various complex games like Starcraft 2 have a strong cognitive component while using neural nets for particular subtasks.

[nl]

I don't think this is convincing.

AlphaGo/MuZero are completely NN based and were so far ahead of the competition when they were developed they led to the whole wave of NN-for-eval that we seen now. And AlphaGo/MuZero doesn't compete in TCEC.

The chess community (especially the stockfish programming group) is very focused on improving their own system. I don't think the fact that is a the strongest system really means much - it's pretty clear they are leaving performance on the table. For example it wasn't until last year (!) that they moved to a GPU based training system.

[Buttons840]

So the answer is probably to find some novel but obvious-only-in-retrospec way to combine neural networks and cognitive methods.

[dqpb]

At the very least, it’s useful for bootstrapping agents.