Comments/Ratings for a Single Item

Thank you Greg, indeed I didn't have it. Now it's alright. Smiling, I say that I will never use a joker. Too much controversial. :=)

Both. You are limited to a maximum of 16 files and 16 ranks. But there is also a limit of 192 squares, so you can't really get 16x16. The largest size is essentially 16x12 or 12x16. Someday this might be increased, but not any time soon. Going to 16x16 would not be too large a code change, but it would have a performance cost for every game, even small ones.

I was thinking of trying some very wide boards, like 8x20 or 8x24, which would fit the 192 squares limit. Is the reason for the 16 file and rank limit that larger dimensions might be more difficult to display clearly?

The way I have always looked at this is that to know whether a square is attacked by the opponent during your turn you should pass that turn, and see whether the opponent then can move there. So the question then is what the turn-pass would do for the movement capabilities of the Joker.

1 seems most natural to me. I would think that passing a turn shouldn't count as moving a piece, so it shouldn't change the Joker's move if it is defined by the last piece moved. In a game that doesn't allow passing, passing should be considered the same as a double move by the other side.

I was thinking of trying some very wide boards, like 8x20 or 8x24, which would fit the 192 squares limit. Is the reason for the 16 file and rank limit that larger dimensions might be more difficult to display clearly?

No, this was not the primary reason.  It has to do with the size of various pre-allocated arrays.  I don't actually remember the implications of more files very clearly ... it might not be that bad.  I will take a look.  Exceeding 192 squares would be far more costly.

2. The turn pass is considered a move of the King

Option 2b: the turn pass can be made by any piece (or any piece w/ a pseudo‐legal move) on the passing side. And is thus, for mate‐detection purposes, equivalent to the intersection of all such moves.

This has the advantage of needing less off‐board state to be maintained: you only need to record what the opponent's last move was, not your own (as with e.g. en‐passant or lion anti‐trading); also a Joker cannot then give check as a piece no longer on the board, which I find a mildly surprising behaviour, and the position in Greg's diagram is unconditionally checkmate. It also matches my proposed update‐on‐touch‐move semantics, which covers castling out of or moving through check, even in the presence of multiple differently‐moving castling‐capable royals, as well as a possible rule for interacting w/ e.g. Orphans

Conversely, opt. 1 has the advantage of being considerably easier (I imagine) to implement, and probably to explain, at the expense of in some ways exhibiting more surprising behaviour. And Daniel's equation of it with a double move makes some sense (though given the context of actual double‐move variants there are possible quibbles). I imagine it'd be the most popular option.

I agree that of the options H.G. listed, 1 seems most natural; I find my opt. 2b a touch moreso, but opt. 1 is not far behind, so people's mileage may (and probably will!) vary

I would only add that H.G.'s #1 can correspond to either my Option 2 or Option 3, depending on how the Joker is defined. (Whether it emulates the last piece moved by the opponent or the last piece moved period.)

Since Aurelian has opted for Option 2, and I'm not hearing any objections (I don't think H.G.'s post really deviated from this), it would be nice to consider the matter settled - at least for Apothecary and for the "default" definition of the Joker. Other game inventors may decided to do different things of course.

EDIT: For the record, these are the potential disadvantages to this option that I've heard: (1) It requires storing state information for the type of piece last moved by both players, and (2) There are circumstances where a Joker can emulate a piece that is no longer on the board (although only when the other side is on the move).

Ok, I think we can forget about 16x16 for now.  I tried ChessV with the 16x12 Double Chess...  OUCH.  It barely works at all.  The branching factor is just too high.  Even the opening position has 72 legal moves (compared to 20 for Chess).  I suspect midgame positions could easily have over 200 although I didn't get that far.

More than 16 files is also problematic.  It would increase the size of a data structure that stores pawn information.

More than 16 files is also problematic. It would increase the size of a data structure that stores pawn information.

Thanks for looking at it anyway

What is the problem with a high branching factor? You can search less deep in a given time, of course, but that also holds for any opponent.

With alpha-beta the EBF only grows as the square root of the typical number of moves. So even if the latter is 6 times higher, the EBF would only go up by a factor 2.5.

What is the problem with a high branching factor? You can search less deep in a given time, of course, but that also holds for any opponent.

I was surprised by just how shocking slow it became.

With alpha-beta the EBF only grows as the square root of the typical number of moves. So even if the latter is 6 times higher, the EBF would only go up by a factor 2.5.

The square root growth is a theoretical value, and while it may be commonly achieved, it is certainly not guaranteed.  What I suspect is happening is this.  The board is very large, the armies start far apart, and no additional evaluation parameters have been added.  So I suspect that there are just a lot of moves where the evaluation is the same so we get way less beta cut-offs.

But it should become easier to find cut moves when the evaluation is constant. If all leaves would be evaluated as 0, the first random sequence of moves will be the PV, and every move you randomly pick in a cut node would immediately be a cut move.

@Greg, If you allow customizable low material evaluation like KMK (M is the CW), may you make it with lists like in KWWK(W is the CF)? I'm sure you though about that, too, and it is a bit late to raise this problem, but it could turn helpfull!

Hello Greg,

Have you an idea about when you will release the next ChessV version?

May you specify if the new version is ready? I am on the tip of my toes for that!

May you specify if the new version is ready? I am on the tip of my toes for that!

I'm sorry, I know you're anxious.  I will post as soon as it is ready, but I can't say when exactly that will be.

Thanks!

What would be good settings for evaluating the balance of a different armies variant? I've tried some tests with 0:30+10 and Medium variation but got conflicting results.

What would be good settings for evaluating the balance of a different armies variant? I've tried some tests with 0:30+10 and Medium variation but got conflicting results.

Your results are understandable.  Although the armies are not balanced, they are close, so you need to play a *LOT* of games to get valid results that are not swallowed up by noise.  At a time control of 30+1, you can't play many games.  And such long controls aren't necessary since with games on an 8x8 board ChessV can reach a search depth of 10 in a fraction of a second on a modern computer.  But the other thing to watch out for is that you aren't playing the same game over and over.  The search variation setting helps, but I also use different pre-calculated opening lines.  I will post more about this soon ...

How many is a lot? I did one set of 200 games and had a final score of -2, but then I tried a second set and by the time it got to 106 games the score was 62.

This is an anomalous result; if the games were really independent, the statistical error should decrease as 1/sqrt(numberOfGames). So for 106 games it would be around 4%, so that the 62/106 (=58.5%) score is about two standard deviations above equality, while the other score pointed to equality with a 3% standard error. The standard error in the differens of the two results should be about 5%, so the 58% is off a bit more than you would expect, but not extremely so.

What I often did to make the games more independent is play them as shuffle games. If you shuffle white and black independently (as seems natural for CwDA) you can create a lot of starting positions even when you leave King and corner pieces in place.

So for 106 games it would be around 4%, so that the 62/106 (=58.5%) score is about two standard deviations above equality

What I meant was that white won 62 games more than black. The actual numbers were 80 white wins, 18 black wins, and 8 draws. This was very different from the first test, where black won more, but with a much smaller margin.

Ah, I see. That is 79%, which is way outside any statistical error bar.

I think I have encountered a new situation involving the joker. I want to post a png with the position but I don't know how. Can someone tell me how to add a picure in the comments?

You just write a FEN of the position between [ fen] and [ /fen] tags, somewhere in the text of your comment.