Accelerating Self-Play Learning in Go · HackerPapers discussions

Accelerating Self-Play Learning in Go — mentioned by carry_bit

Thu, 03 Nov 2022 17:30:07 GMT

From a tweet of one of the authors [1]: "KataGo was trained on Tromp-Taylor rules so we evaluate our attack using this too." This is incorrect. According to the KataGo paper [2], KataGo is trained using modified Tromp-Taylor rules (as pointed out by [3]): "Self play games used Tromp-Taylor rules modified to not require capturing stones within pass-aliveterritory [...] In Go, a version of Benson's algorithm [1] can prove areas safe even given unboundedly many consecutive opponent moves ("pass-alive"), enabling this minor optimization." It would be more interesting if they beat KataGo using the…

Paper and discussion history. A mention is not an endorsement.

Accelerating Self-Play Learning in Go — mentioned by cgreerrun

Mon, 22 Jun 2020 14:54:13 GMT

I've been working on a Python implementation that uses Gradient Boosted Decision Trees (LightGBM/Treelite) instead of using a neural network for the value/policy models: https://github.com/cgreer/alpha-zero-boosted It's mostly to understand how AlphaZero&Friends work. I'm also curious about how well a GBDT could do, and if there are self-play techniques that can accelerate training. The nice thing about a GBDT is that, unlike when using a NN, you can do thousands of value/policy lookups per second on a single core. So it should be cheaper to scale self-play and run a lot of self-play…

Paper and discussion history. A mention is not an endorsement.

Accelerating Self-Play Learning in Go — mentioned by jedharris

Sun, 14 Jun 2020 22:16:07 GMT

Related demos, repositories, papers and posts: More recent and extensive post: https://blog.janestreet.com/deep-learning-the-hardest-go-pro... Site to play against KataGo: https://online-go.com/player/592684/ Arxiv paper on KataGo design: https://arxiv.org/abs/1902.10565 GitHub repository of KataGo: https://github.com/lightvector/KataGo Another writer commenting on the history of Go AI, and significance of KataGo: https://brantondemoss.com/writing/kata.html

Paper and discussion history. A mention is not an endorsement.

Accelerating Self-Play Learning in Go — mentioned by sanxiyn

Thu, 11 Jun 2020 10:07:56 GMT

KataGo optimized AlphaZero and achieved 50x compute reduction. https://arxiv.org/abs/1902.10565

Paper and discussion history. A mention is not an endorsement.

Accelerating Self-Play Learning in Go — mentioned by gwern

Sun, 01 Dec 2019 01:27:42 GMT

He's also wrong about a lot. For all his insistence on accuracy, he himself is misleading or ignorant. Like his slam of the Dartmouth project - even if you knew nothing about it, all you have to do is click through to see the claims of 'solving vision' are sheer projection and urban legend. And he's happy to make up claims out of whole cloth: for example, when he says "AlphaGo works fine on a 19x19 board, but would need to be retrained to play on a rectangular board; the lack of transfer is telling.", for which he provides precisely zero evidence, he ignores the fact that AG training works…

Paper and discussion history. A mention is not an endorsement.

Accelerating Self-Play Learning in Go — mentioned by codehotter

Fri, 22 Nov 2019 20:16:28 GMT

AlphaGo Master, unsurprisingly, was significantly stronger than AlphaGoZero. AlphaZero, although it can play multiple games, was weaker yet. In both cases, they compared the 40 block version of the one with the 20 block version of the other (they had to double the network size to approach the level of the predecessor.) Recently, Katago has reached similar levels of strength using a small fraction of the resources: https://arxiv.org/abs/1902.10565 It depends on what you mean by "more efficient." The significance of AlphaZero was that you can reach good results in a variety of domains even…

Paper and discussion history. A mention is not an endorsement.