The basics of study with Katago concern calibration with komi, numbers of playouts, precision of values and study modes.

KOMI

For every ruleset, there is a komi so that initially the winning probability is about 50.0% and the expected score is about 0.0. One can set such a komi or accept other initial values. Later values must be compared to the initial values.

NUMBERS OF PLAYOUTS

Usually, Katago frequently changes its mind about the relative orders of top move candidates if the number of playouts per top move candidate is smaller than 100,000. Therefore, usually, one must not trust Katago's choice on the next move in every particular current position if the number of playouts per top move candidate is smaller than 100,000. Exceptionally, there can be only one urgent move candidate and only then a smaller of playouts is sufficient. If, at 100,000 playouts per top move candidate, two or several of them have very close values, reversion might be the cause, one can accept them as essentially equal or wait for a higher (if necessary, much higher) number of playouts per top move candidate until values differ and the relative order of top move candidates stabilises. If the GPU is too slow for at least 100,000 playouts per top move candidate or the GUI hides numbers of playouts, do not use Katago for best next move predictions but only use it to play against it or identify big blunders!

PRECISION

Typical GUIs show values with either a 0.1 or 1 point precision. For the purpose of seeing big blunders, the precision 1 point is sufficient and kyu players might focus on this first. For the purpose of seeing the best next moves, the precision needs to be 0.1 point. If instead one sets the precision 1 point or uses a GUI that only offers this, the error of the expected score is 0 to 1 so on average 0.5 per move. Since 14 points equals one rank, one can lose one rank every 28 moves. For dan players or kyus wishing to reach dan level, this is inappropriate. They should also recognise and learn from the small mistakes so use the precision 0.1 point.

STUDY MODES

Katago allows the following study modes:

- Permanent pondering: Katago always thinks, the user chooses or undoes moves, or walks through the variations.
- Analysis: Katago analyses, afterwards the user views the findings.
- Playing: The user plays against Katago.
- Katago's self-play with one or two engines.

GUIs might offer further variations on the modes.

Is there any guidance on studying individual joseki - for example how to set out the rest of the board in a relatively "neutral" position that allows study to be focused on local joseki variations ?

For this purpose, start with one joseki corner and three single stone corners. To limit the amount of study, occupy the diagonally opposite corner with a 4-4 stone. In the two adjacent corners, each choose either a black or a white stone. Either such stone is 3-4 or 4-4 (because mostly others are inferior in AI play, with a few exceptions). Either 3-4 stone vary as open or closed to the adjacent joseki corner. Maintain alternation on the whole board. Now, study the joseki one position / next move to be found at a time and eventually explore variations depending on different environments.

Be aware that tenukis and later joseki continuations or follow-ups can occur. For special purposes, such as ladder conditions, you might need to alter the environment suitably. For the early middle game, joseki play can differ so extended study also uses middle game environments.

Is it correct to consider "best move in terms of unlimited time/playouts (theoretical best move) or to consider "best move" in terms of best move identified under the practical limitation of time/playouts" (practical best move). I think we need to consider what is happening when move A is evaluated as better than move B given the practical limitation but move B ends up evaluated better than move A when that greater (impractical under time control) number of playouts is used. Doesn't that mean that there is some obscure, hard to discover "refutation" of some lines?

To me at least, it makes sense to say A is the best move under time control X but B is the best move under time control Y. In other words, even if we knew knew B, it might not be better than A if under time control X. We might not only need to be under Y to identify B, but also to find the right subsequent moves that made it better than A (in the absolute sense)

Neural net computer go is not mathematical theorem proving, that is, we do not know whether a move is perfect play. Some practical search limit is needed, although one can fill one's RAM if a particular query is important. Go tournament players' thinking time are a different consideration and simplifying but suboptimal moves can be more suitable, unless players study the complications carefully to be well prepared for tournament play. For the AI researcher, both the possibly more complicated but correct move and the simpler but possibly suboptimal move are interesting besides explanations why one is better and observations how much the other loses.

Only few lines are obscure or rely on spectacular tesujis. From a high amateur dan player's view, AI plays surprisingly many ordinary moves. It is just that there are many such moves and the right timing and choice fitting the environment creates the particularly good results.

Agreed in general BUT we were not talking about moves in general. MOST of the time we do not have the situation where A>B at a lower number of playouts but B>A at a very high number of playouts. Nor is it likely that one of the FIRST moves (A or B) is a tesugi. More likely tat any tesugi/obscure move involved is several moves down the line, and THAT is why it took a larger number of playouts to discover the line that caused the switch between A and B.

It is my observation in Katago study that it changes the relative order of most move candidates frequently for low numbers of playouts. A>B and B>A tend to change several times before either order stabilises for high numbers of playouts.



AI and humans find it easier to recognise or assess any tesuji if it is a possible next move in the current position. Regardless, if such a tesuji is uncommon, also Katago can have its difficulties to choose it and many playouts can be needed.

With the following assumptions, however, your reasoning becomes more realistic:

- There have already been at least 100,000 playouts per top move candidate in the current position.

- Exactly two top move candidates are clearly better than all other move candidates.

- These same two top move candidates remain clearly better than all other move candidates for increasingly higher numbers of playouts.

- The two top move candidates do not start reversions, which AI deems correct, to any same follow-up position and are not accidentally otherwise equally valuable.

Then your statement is correct. Most of the time (that is, more often than not), the relative order A>B at 100,000 playouts for each of A and B will be the same as the relative order for high numbers of playouts. Even so, between 100,000 playouts and high numbers of playouts, the relative order can temporarily change to B>A once or several times.