Collection of Instructive Havannah games?

In the Castro thread we are discussing one

specific bot weakness, shown in a very instructive

game, lost by the bot.

Proposal: to make a collection of instructive

Havannah games for bot programmers. (Typically

most of these games will be games lost by a bot.)

Ingo.

Yes Ingo, that's a nice idea. I am using something like that for R-o-F testing:

e.g., the engine should find the following moves:

lg 1242741 : move 12 is A5 ;

lg 1170111 : move 30 is C2 ;

lg 1165916 : move 34 is M6 ;

lg 1241219 : move 14 is J3 ;

lg 1241197 : move 7 is B3 ;

lg 1240767 : move 19 is B4 ;

lg 1239553 : move 9 is G2 ;

lg 1239472 : move 13 is G2 ;

lg 1242768 : move 39 is E4 ;

lg 1242134 : move 9 is F5 ;

but in fact currently R-o-F gets at most two moves right from that list.

Thomas warns me that what might seem obvious to a human,

may not be the only move, or may even be totally wrong.

So one should perhaps not just test whether the engine

finds the move from the test spec, but also (in case it suggests another)

how the evaluation for the intended move actually was.

Have you tested these position with really long evaluations? Let it run on those positions for a couple hours and see if it gets the right answer, or maybe it can give you a better answer. Then shorten the time to see how little time you need to get the right answer. Are any of these late enough in a game that you can actually prove the position?

@castro:

1. yes, I know what I could do with these if I had the time.

Ingo asked for a collection of test cases, and I provided some.

2. "prove the position": mostly not, in the sense of a complete (enough)

game tree with a definite plus/minus infinite root labelling.

But that's not the point (expect for the few endgame problems):

E.g. in this position White 12. A5 is absolutely required to prevent black's fork. After that it's still difficult (Black 13. I7) but if black gets A5 it's just hopeless. So, coming back to Ingo's question,

this test case is instructive because it would show whether the player knows about knight's moves and forks.

From the "better late than never" front, I also agree that there is benefit from having a nice test suite. I've only looked at the first game from Johannes' list (#1242741) but Wanderer takes far too long to find A5: over 3 minutes. I will try to add some positions I feel are critical shortly. There are certainly plenty to choose from from Wanderer's losses. For example, there are many losses that have this form and this needs to be addressed.

[game;id:1320115;move:19;title:Why doesn't Wanderer see the fork is hopeless here?]