3 ms·

I've been working on a similar idea of learning a latent vector representation of chess positions and moves. I'm not sure whether to be happy or disappointed th

I've been working on a similar idea of learning a latent vector representation of chess positions and moves. I'm not sure whether to be happy or disappointed that this paper is much more simplistic than what I've been trying to do. They call it "machine learning", but they're just doing PCA and nonnegative matrix factorization, neither of which falls into the class of methods that most people think about when you say "machine learning." Their proposed prediction method will often (perhaps usually) predict illegal moves--it might propose moving a piece that isn't even on the board, for example. Understandably they don't show any examples of actually applying it.

nestorD 7y ago
Well, word2vec has been reduced down to matrix factorization with Glove (the reformulation into linear algebra managed to get similar result but one order of magnitude faster) so it does make sense to try that first.
Der_Einzige 7y ago
Dimensionality reduction is absolutely a form of "machine learning". UMAP or Ivis are doing the same thing but utilize fancier methods so now they must be more relavent? I'd be curious to see what happens when you combine the system with a system that removes the possibility of illegal moves. I bet accuracy will go dramatically up - especially if they utilize a really modern transformers based architecture for the vectorization part.
- CrazyStat 7y ago
  Dimensionality reduction more generally may well be a form of machine learning. Doing really simple dimensionality reduction and then a really stupid prediction method doesn't quite cut it as far as I'm concerned.
  - jmmcd 7y ago
    > doesn't quite cut it I think the underlying assumption here is that some approaches are too simple to "cut it", ie to be called machine learning. And eg PCA is too simple, and so should be excluded, and perhaps categorised as "just" statistics or linear algebra depending on your point of view. But in my view, it's better to let go of this view which implicitly hypes ML in a damaging way. F = ma is still physics, even if simple. k-NN and PCA are still ML.
    - CrazyStat 7y ago
      Eh, I disagree. Not from a normative point of view, but from a descriptivist one. Most people don't consider doing a PCA to be machine learning. I'm a statistician, not a machine learning guy. PCA is a great statistical method. But doing a PCA and calling it machine learning seems a tad misleading.
      - jmmcd 7y ago
        Well, I certainly agree that many people use ML in this way. But to me it's just bad terminology. The thing that happens during ML -- the L part -- must be the same as what happens during statistical learning. Many (sophisticated) ML algorithms are statistical in nature, so there is nothing there that stops simple stat-learning algorithms from being ML algorithms. And there is nothing in the terminology that says ML has to be sophisticated relative to stat learning.

nestorD 7y ago

Well, word2vec has been reduced down to matrix factorization with Glove (the reformulation into linear algebra managed to get similar result but one order of magnitude faster) so it does make sense to try that first.

Der_Einzige 7y ago

Dimensionality reduction is absolutely a form of "machine learning". UMAP or Ivis are doing the same thing but utilize fancier methods so now they must be more relavent? I'd be curious to see what happens when you combine the system with a system that removes the possibility of illegal moves. I bet accuracy will go dramatically up - especially if they utilize a really modern transformers based architecture for the vectorization part.

CrazyStat 7y ago
Dimensionality reduction more generally may well be a form of machine learning. Doing really simple dimensionality reduction and then a really stupid prediction method doesn't quite cut it as far as I'm concerned.
- jmmcd 7y ago
  > doesn't quite cut it I think the underlying assumption here is that some approaches are too simple to "cut it", ie to be called machine learning. And eg PCA is too simple, and so should be excluded, and perhaps categorised as "just" statistics or linear algebra depending on your point of view. But in my view, it's better to let go of this view which implicitly hypes ML in a damaging way. F = ma is still physics, even if simple. k-NN and PCA are still ML.
  - CrazyStat 7y ago
    Eh, I disagree. Not from a normative point of view, but from a descriptivist one. Most people don't consider doing a PCA to be machine learning. I'm a statistician, not a machine learning guy. PCA is a great statistical method. But doing a PCA and calling it machine learning seems a tad misleading.
    - jmmcd 7y ago
      Well, I certainly agree that many people use ML in this way. But to me it's just bad terminology. The thing that happens during ML -- the L part -- must be the same as what happens during statistical learning. Many (sophisticated) ML algorithms are statistical in nature, so there is nothing there that stops simple stat-learning algorithms from being ML algorithms. And there is nothing in the terminology that says ML has to be sophisticated relative to stat learning.