ELF OpenGo-level 19x19 Go from scratch at one tenth of KataGo's reported compute (about 51 V100-days)
Difficulty 35: KataGo already achieved a 50x reduction in 2019, and later search improvements (Gumbel planning) suggest a further 10x is attainable but undemonstrated; mostly engineering and compute. Application value 12 and leverage 12: efficiency gains may transfer to other planning domains, but Go itself has no further practical stake.
The problem
Build a world-class Go player from scratch using one tenth of the computing a leading open project reported.
- 1Self-play generates training games
- 2Search improves the network's guesses
- 3Network trains on search results
- 4Cut cost with better targets
- 5Match reference in 400 games
Why it matters. AlphaZero learned superhuman Go from self-play alone, but with thousands of TPUs. KataGo (Wu, 2019) showed that general changes to the recipe cut the computation by about 50 times. Another tenfold cut would put strong Go training within reach of a university lab, and would test whether the efficiency ideas are general enough to carry over to other planning problems.
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Step 2
- Policy gradient methods and PPOAlgorithm
- Step 3
- Monte Carlo tree search and AlphaZeroAlgorithm · host
- Then
- ELF OpenGo-level 19x19 Go from scratch at one tenth of KataGo's reported compute (about 51 V100-days)The problem
State of the art
KataGo (Wu, 2019, arXiv) reports a 50-fold computation reduction over comparable methods (AlphaZero, ELF OpenGo, Leela Zero) through mostly non-domain-specific changes (auxiliary policy and ownership targets, playout-cap randomisation, global pooling), and surpassing ELF's final model after 19 days on about 27 V100 GPUs (a maximum of 28 at any time), about 1.4 GPU-years against ELF's roughly 74 GPU-years. Gumbel AlphaZero/MuZero (Danihelka et al., ICLR 2022) improves policy learning with few simulations and is implemented in mctx. Later KataGo training runs used further improvements but did not publish a clean compute-to-ELF-strength figure under the same protocol, and no independently reproduced run reaching ELF strength at one tenth of the 2019 compute has been published.
Realistic applications
Cheap superhuman self-play would make search-plus-learning research accessible and tests whether efficiency tricks transfer to other planning domains.
What counts as solved
Released code and training logs, reproduced by an independent group, showing the strength match within the V100-equivalent compute budget.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
Task: 19x19 Go under stated rules, trained from random initialisation without human games. Metric: score of the final network against the released final ELF OpenGo model (and, as a second anchor, a fixed public KataGo network) at equal visits per move, at least 400 games, with a binomial 95% confidence interval; pass if the point estimate is at least 50% and the lower bound at least 45%. Compute: at most 51 V100-equivalent GPU-days counting self-play, training and gating, logged with wall-clock timestamps; for non-V100 hardware, report the throughput ratio of the same KataGo benchmark command (fixed network, visits per second) measured on both GPU types and convert. Seeds: two full runs if affordable, otherwise one run plus a bootstrap over evaluation games, stated clearly.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Engineering and compute; improvements are mostly empirical efficiency gains.
Bottlenecks
ranked by votes- #1ComputeSelf-play data generation
Most compute goes to generating self-play games; at about 51 V100-days the budget allows roughly a tenth of KataGo's 4.2 million 2019 games, so each game must carry more training signal.
Community grades
Sign in to add your own grade.
Connected problems
No connections yet.
Progress
- Feb 27, 2019KataGosource
Reports surpassing ELF OpenGo's final model after 19 days on about 27 V100 GPUs, a 50x compute reduction.