3 ms·
From an offline chat with the original author, The ELF OpenGo paper[1], which is an open implementation of AlphaGo Zero developed by Facebook AI: "First, we t
by newswasboring 6y ago
From an offline chat with the original author,
The ELF OpenGo paper[1], which is an open implementation of AlphaGo Zero developed by Facebook AI:
"First, we train a superhuman model for ELF OpenGo. Af-ter running our AlphaZero-style training software on 2,000GPUs for 9 days, our 20-block model has achieved super-human performance that is arguably comparable to the 20-block models described in Silver et al. (2017) and Silveret al. (2018)."
[1]: https://arxiv.org/pdf/1902.04522.pdf https://arxiv.org/pdf/1902.04522.pdf