KataGo Opening Books

Here are opening books computed by KataGo for small boards! These books have full symmetry and transposition handling, including ko and superko-aware handling of transpositions, and as of 2023 hopefully are the most thorough opening analyses for these board sizes published so far.

For smaller boards like 7x7 the books are possibly approaching optimal, and on 8x8 might be accurately enumerating the majority of optimal opening lines even if many less important lines are uncertain. For 9x9, analysis remains much harder, but hopefully those books still give high-quality and mostly accurate coverage of a wide variety of openings.

Update November 2024:
Added more variations to 9x9 books, including analyzing many deeper openings common in strong GoQuest games, 3-3 invasions under the 4-4 point, and various sharpening of many side lines. The old books are also still there so you can compare.
Update October 2023:
9x9 books are now posted below!
Also you can download the strong finetuned 9x9 neural net used to build those books, download here!
And see here for some high-level notes on the book and interesting moves!

Thanks again to everyone who has made KataGo and this research possible, and especially to those people who provided feedback on this work!

9x9 Book, Tromp-Taylor Rules, Komi 7 (last updated 2024-11-05)     Download for viewing offline    Old TT book 2023-10-27    Old TT download 2023-10-27
9x9 Book, Japanese-like Rules, Komi 6 (last updated 2024-11-05)     Download for viewing offline    Old JP book 2023-10-27    Old JP download 2023-10-27
8x8 Book, Tromp-Taylor Rules, Komi 10 (last updated 2021-11-14)     Download for viewing offline
8x8 Book, Japanese-like Rules, Komi 9 (last updated 2021-11-14)     Download for viewing offline
7x7 Book, Tromp-Taylor Rules, Komi 9 (last updated 2021-08-06)     Download for viewing offline
7x7 Book, Japanese-like Rules, Komi 8 (last updated 2021-08-06)     Download for viewing offline

Additionally, thanks to an excellent contributor Friday9i who worked on book generation and a fine-tuned neural net for smaller boards, here are some additional books. Some books for additional small sizes may be added later too. Some of these books may be expanded a bit less thoroughly than the above books.

If you enjoy this and would like to support KataGo's improvement and research, KataGo is continuing to train at katagotraining.org via volunteer crowdsourced computation. The number of contributors is small enough that additional people and GPUs can make a significant difference - if you are interested, please consider helping out!

For Questions

If you have questions about these books, interesting discoveries, refutations of the moves, questions about KataGo or how to help contribute, or anything else to ask or share - please feel free to drop in to Discord chat, leave a message/issue on GitHub, or otherwise get in touch - would love to hear more.

Rules and Komi Used

Tromp-Taylor rules: These are strategically the same as Chinese rules, except for using positional superko instead of simple ko, and a few rules beasts like "1-eye-flaw". It is rare for naturally-occurring positions to depend on the choice of ko/superko rule or to be affected by such differences, so for most practical purposes these books are Chinese rules opening books too.

Japanese-like rules: Although the full Japanese rules are difficult to implement for computer programs, KataGo achieves a reasonably close match, including handling of points in seki, bent-four-in-the-corner, needing to protect a final ko, etc. One mild difference is that passing without filling all dame is permissible in KataGo's rules (as on many online servers). This almost never affects the strategic outcome. Aside from that, a few genuine strategic differences are known but they are uncommon.

Komi: For each ruleset, we use the komi believed to be optimal for that size.

Optimal komi for 7x7 appears to be 9 in area-scoring rules like Tromp-Taylor or Chinese rules, and 8 in Japanese scoring rules, and from smaller experiments also does not seem to depend on things like the choice of ko or superko rule.
Optimal komi for 8x8 similarly appears to be 10 in area-scoring rules and 9 in Japanese scoring rules.
Optimal komi for 9x9 appears to be 7 in area-scoring rules and 6 in Japanese scoring rules.

Jump To Contents

Is 7x7 Go Solved?
How About 8x8 Go?
How About 9x9 Go?
How These Books Were Generated
7x7 Highlights and Discoveries
8x8 Highlights and Discoveries
9x9 Highlights and Discoveries

Is 7x7 Go Solved?

In a rigorous mathematical sense, the above books do NOT "solve" 7x7 Go, since even if the main lines luckily all turn out to be exactly optimal, there is not a proof of this, and the books are of course astronomically far from being exhaustive.

There are almost certainly multiple mistaken evaluations in the book and/or suboptimal move choices. One can find many positions where KataGo itself indicates obvious uncertainty, such as reporting winrates that are far from any of 0%, 50%, or 100%.

However, despite no rigorous proof, we do suspect that most of the preferred lines of the 7x7 books are optimal, maybe even almost all.

Surprisingly, at basically no point during the book-building process did KataGo substantially change its mind about any of the main lines branching off in the first 5-7 moves. Even as the book grew 5x, 10x, 50x, the results have been very stable. The various mistakes and blind spots that the book *has* found and corrected over real-time KataGo have so far been limited to variations not affecting the top-level results, such as side variations after mistakes or dispreferred moves by both sides, or deep variations within a main line for which there were other working alternatives.

With an exponentially large space it is hard to eliminate all mistakes, and there is still some chance of a significant mistake somewhere, but we might not be too far from the point where a bot that always played according to the preferred move from the book until reaching a leaf, and then finished the game with a sufficiently powerful search at runtime, would be literally unbeatable.

If you find any large mistakes in these books, particularly ones that could affect the value of the main lines or immediate side variations, that would be fascinating material to study, please let us know!

Note that all scores in the books should be considered approximate. Scores should be a decently good prediction for the true score when the difference is not too large, but they are NOT expected to always be perfect. They are very often not almost-an-integer, reflecting this uncertainty, and it is entirely expected that occasionally moves might only be win-loss-draw-optimal, rather than score-optimal. Despite this, KataGo appears to do a good job differentiating between moves of different scores anyways, including finding many 1-and-2-point improvements over earlier human-driven 7x7 solution attempts.

How About 8x8 Go?

On 8x8, the number of optimal lines seems to be vastly larger than on 7x7, and correspondingly the confidence with which we are able to enumerate these lines is much lower than. Unlike 7x7, where almost every optimal opening line has been seemingly identified and explored deeply, and almost every deviation at any point is confidently rejected, on 8x8 it is not hard to find variations that KataGo is *not* confident in, particularly when repeatedly following moves that are evaluated as likely optimal but are second, third, fourth, etc choice, or moves that are unusual and "bad shape" and require knife-edge precision followups to justify why they still barely work.

Similarly when generating the 8x8 books, among the hundreds of important variations in the first 5-10 ply, several times KataGo changed its mind on some of these variations, discovering a believed-losing move was actually also optimal, or discovering that an apparent drawing line with 50% winrate and equal score after millions of visits was instead a loss. A small fraction of branches may still be wrong in such a manner. By contrast, this almost never happened on 7x7.

Despite this, the large majority of lines in the 8x8 books are still probably correct, particularly the lines that follow mostly the first-choice moves and/or moves with high policy prior and natural good shapes. This is particularly the case for the 8x8 Japanese rules book. Due to the finer scoring granularity of Japanese scoring over standard area scoring, there are far fewer optimal lines under Japanese rules, making it much easier to validate them thoroughly. For most lines, the 8x8 Japanese rules book may be not far from the confidence of the 7x7 Tromp-Taylor book at being truly the optimal solution.

How About 9x9 Go?

Obviously, for 9x9 Go the books are even much more uncertain than for 8x8.

For small boards, the books on this site might sometimes be considered a non-rigorous practical "solution" for the game. However, for 9x9, even as early as move 2 there are moves whose winrates are not close to any of 0%, 50% or 100%, indicating plenty of uncertainty in the ultimate outcome. So the book really is just a compilation of the (hopefully still high-quality) opinions of a top-level bot at very high playouts, rather than anything resembling a "solution".

There are almost certainly some deep misevaluations and mistakes in the 9x9 books, due to the sheer exponential number of branches. If you find any particularly big mistakes, particularly that affect evaluations of major opening lines, feel free to reach out the Computer Go discord or open an issue at KataGo's GitHub page. It is often possible to fix mistakes by selectively expanding relevant branches.

How These Books Were Generated

These books were generated by incrementally growing a tree from the opening position of each board over repeated iterations, in each iteration:

Choosing up to K nodes in the tree to expand, based on stats at each node about the possible moves that are not yet part of the tree.
For each chosen node in parallel, running a full search to find KataGo's most-preferred move to add to the tree. Moves already in the tree are banned and treated as illegal by the search.
Then running a quicker search with the new added move also banned, to update the stats about the remaining moves not yet part of the tree.

For example, for the 7x7 books, KataGo generally used about 15000 visits for full search, and about 7500 visits for the quicker search, with wide root noise of 0.10. KataGo performed many thousands of iterations, with K=64 parallel nodes expanded per iteration. Additionally, the 7x7 books for the two rulesets were used to seed each other a few times, each one forced to expand every node the other did, giving much more exploration to the tiny one-point differences that make the difference between a win/loss under each.

Overall, the above method of expanding the tree appears to greatly reduce inaccuracies relative to real-time KataGo. Always forcing the search, after picking out its favorite move, to invest many visits into the next best moves while banning the more favored moves, forces exploration of moves that would normally be searched only shallowly. This prevents surprising correct lines or refutations from being overlooked. Within the horizon of the book itself, it should do a very good job fixing errors.

Choosing Nodes to Expand

The way the nodes to expand on each iteration were chosen was by:

Assigning a cost to each edge of the tree.
Computing the path cost of a node to be the minimum sum of costs needed to reach that node from the root.
Adding a node-specific additional cost to each node's path cost, to get a final cost.
Selecting the K nodes with the smallest final cost.

Costs were based on a variety of factors, including the raw policy prior of moves, the current winrate and score of the moves based on the tree so far, the number of moves expanded at a node already, and some crude measures of uncertainty. These were adjusted to give a good balance between getting enough diversity of moves explored and exploring the most important lines deeply.

Many manual adjustments to costs were also added to force much deeper analysis of most lines claimed to be optimal by earlier human-published work, or that involved difficult refutations of suboptimal moves, some of which are discussed further down, and especially where KataGo's book did not appear to expand such lines very deeply on its own.

For 7x7, surprisingly, and promisingly as a sign of the accuracy of those books, there does not seem to have been any case so far where this deeper forced analysis overturned KataGo's initial judgment and revealed that a rejected line was an improvement over the preferred line. And in fact, it was hardly ever observed that KataGo even on its own changed its shallower judgment with a deeper book expansion on any important lines. However, manual deepening it did help sharpen the evaluations of a huge number of side variations (e.g. sharpening a 30% winrate move towards being either a loss or being equally good as the preferred 50% winrate move), and help grow the book deep enough to resolve many tactics in those side variations where KataGo did still have nontrivial blind spots for the book algorithm to correct even though they didn't affect the top-level result.

For 8x8 and especially 9x9 books however, this has not been the case. For these larger boards, KataGo definitely has both its own and with targeted manual help significantly changed or overturned its opinion of an opening. Larger boards are hard!

Transpositions and Symmetries

KataGo's book has full symmetry-aware superko-aware transposition handling, using a 256 bit hash.

We handle symmetries by hashing the board once for each of the possible 8 orientations and taking the minimum to be the canonical hash. This canonical hash is used to identify and share the same node between different orientations of a position and between different transposed move orderings that reach that position. This includes even move orderings that transpose not to the same position but rather to different orientations of the same position.

Edges of the tree keep track of the dihedral symmetry transformation that relates their two endpoints. These transformations are successively composed as you walk the tree, so that the board always is presented in the orientation that is consistent with the path by which you reach a particular node, regardless of what the canonical orientation of that node is.

The hash also takes into account a heuristic amount of the history, such that enough of the history is preserved to maintain full superko correctness despite the fact that transpositions are allowed. The heuristic is that the history is cleared any time a move is made such that S + E > 9, where S is the total number of stones in the resulting chain containing the stone and E is the total number of empty spaces in all empty regions bordering that chain (e.g. nearly the whole empty area of the board at the start of the game). This is easily proven to be a lower bound on the length of any possible superko cycle involving that move, so for this heuristic to fail, it requires a superko cycle be quite long and involve the repeated capture of decently large groups (e.g at least 4-5 stones). If there is a beast where a line necessary for correct play violates this bound, I would be interested to know about it.

Newer details for 9x9 Book Generation

For the 9x9 books, KataGo spent a few months in early 2023 making a first attempt to generate 9x9 books. Eventually it became clear that despite being among the strongest bots on 9x9, the evaluation quality still left a lot to be desired when trying to solve many important variations in 9x9.

Therefore, during mid-2023, KataGo spent about 3 months finetuning a 9x9-specific neural net on a few strong home GPUs. Surprisingly, this was enough. Despite having far less compute power than the full-size training fueled by many generous contributors at katagotraining.org, the 9x9 finetuning run worked both because it was able to start from the already well-trained full-size neural nets, and because training on 9x9 is vastly cheaper. The games last fewer moves, on each move the same number of playouts reaches deeper and solves more tactics accurately, for each playout the neural net query is much cheaper due to the smaller board, and so on.

As a result, it was possible to obtain a neural net far stronger on 9x9 than KataGo's normal nets and that better handled most of the variations that gave trouble to the older net. This new neural net was used to generate the 9x9 books. The net is also available for download here!

Aside from the new neural net, the 9x9 books used a similar approach to 7x7 to expand the tree, except with some different choices of parameters. Most notably, book expansion used up to 500000 visits at a time instead of 15000, but added the whole tree of nodes to the book from each such subject to those nodes having enough visits rather than adding just one node. The 9x9 books also enforced at least 10000 visits per node, and followed the same method of re-searching remaining moves while banning more favored moves. This continued to be extremely effective at surfacing and correcting blind spots within the horizon of the book.

7x7 Highlights and Discoveries

See this page for an overview of some of the most fascinating and fun variations and discoveries from KataGo's 7x7 books, relative to earlier attempts at solving 7x7!

8x8 Highlights and Discoveries

See this page for an overview of the main seemingly-optimal variations for 8x8 and a sampling of some of the interesting moves!

9x9 Highlights and Discoveries

See this page for an overview of the book for 9x9 and a sampling of some of the interesting moves!