7 ms·

I love the idea of cognitive AI. I've dabbled with OpenCog https://en.wikipedia.org/wiki/OpenCog https://en.wikipedia.org/wiki/OpenCog , and in my youth I read

I love the idea of cognitive AI. I've dabbled with OpenCog https://en.wikipedia.org/wiki/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 4y ago
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.
- Buttons840 4y ago
  So the answer is probably to find some novel but obvious-only-in-retrospec way to combine neural networks and cognitive methods.
- nl 4y ago
  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.
  - mtlmtlmtlmtl 4y ago
    AlphaGo/MuZero/LeelaZero are not completely NN-based. They're still Monte Carlo Tree Search. Completely NN would have to be some sort of bitboards in moves out thing. That's not a thing afaik. They rely a lot more on the NN though is my understanding, sacrificing search depth for the compute. LeelaZero does compete, and Stockfish tends to outcompete it. It's my feeling that MCTS, while necessary with go's branching factor making minimax approaches untenable, is not actually a win in chess compared to alphabeta paired with the decades of heuristics that have been developed. I have some complex reasons for believing this. One of them is that Stockfish' move ordering relies heavily on statistical "learning"(histories, continuations, killer move heuristic, etc) heuristics that are probably more powerful if your search is wider. This probably applies to Transposition/pawn tables as well. More importantly, I'm sceptical of the amount of useful inference that can be made from a static position in chess vs go. I barely even know the rules of go, but I am an expert level chess player. I think go has a lot more static structure that likely makes larger NNs more useful than they are in chess. Chess is far more illuminated by specific tactical themes in the position that are very effectively picked out through qsearch paired with move ordering heuristics and caching. With chess, I think there's some point when inferring anything more is just gonna involve search anyway. And you don't want ad hoc search appearing in the eval function because that won't benefit from tables and move ordering heuristics. And yes, certainly there's lots more to gain for Stockfish, and their incessant improvement is impressive. And obviously the size of their community is a huge advantage. But I don't actually think MCTS is the way to go for chess, ultimately. Minimax based techniques are just too damn effective, with a sprinkle of NN. Shogi is an interesting in between here. Conceptually like chess, but with a much higher branching factor. I'm not at all aware of the state of the art of shogi minimax search though.
    - evolvingstuff 4y ago
      So I just googled "differentiable Monte Carlo" and it looks like that concept has so far only been applied to ray tracing, but I would be shocked if that (or similar) doesn't become a thing in the next year or three, such that AlphaGo et al become end-to-end differentiable.
dqpb 4y ago
At the very least, it’s useful for bootstrapping agents.

mtlmtlmtlmtl 4y ago

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.

Buttons840 4y ago
So the answer is probably to find some novel but obvious-only-in-retrospec way to combine neural networks and cognitive methods.
nl 4y ago
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.
- mtlmtlmtlmtl 4y ago
  AlphaGo/MuZero/LeelaZero are not completely NN-based. They're still Monte Carlo Tree Search. Completely NN would have to be some sort of bitboards in moves out thing. That's not a thing afaik. They rely a lot more on the NN though is my understanding, sacrificing search depth for the compute. LeelaZero does compete, and Stockfish tends to outcompete it. It's my feeling that MCTS, while necessary with go's branching factor making minimax approaches untenable, is not actually a win in chess compared to alphabeta paired with the decades of heuristics that have been developed. I have some complex reasons for believing this. One of them is that Stockfish' move ordering relies heavily on statistical "learning"(histories, continuations, killer move heuristic, etc) heuristics that are probably more powerful if your search is wider. This probably applies to Transposition/pawn tables as well. More importantly, I'm sceptical of the amount of useful inference that can be made from a static position in chess vs go. I barely even know the rules of go, but I am an expert level chess player. I think go has a lot more static structure that likely makes larger NNs more useful than they are in chess. Chess is far more illuminated by specific tactical themes in the position that are very effectively picked out through qsearch paired with move ordering heuristics and caching. With chess, I think there's some point when inferring anything more is just gonna involve search anyway. And you don't want ad hoc search appearing in the eval function because that won't benefit from tables and move ordering heuristics. And yes, certainly there's lots more to gain for Stockfish, and their incessant improvement is impressive. And obviously the size of their community is a huge advantage. But I don't actually think MCTS is the way to go for chess, ultimately. Minimax based techniques are just too damn effective, with a sprinkle of NN. Shogi is an interesting in between here. Conceptually like chess, but with a much higher branching factor. I'm not at all aware of the state of the art of shogi minimax search though.
  - evolvingstuff 4y ago
    So I just googled "differentiable Monte Carlo" and it looks like that concept has so far only been applied to ray tracing, but I would be shocked if that (or similar) doesn't become a thing in the next year or three, such that AlphaGo et al become end-to-end differentiable.

dqpb 4y ago

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