Fortress Draws

[AdminX]

If Computer Chess Programs cannot be taught to recognize Fortress Draws, could not a Fortress Tablebase be created to overcome this problem? Of I know it would always be a on going work in process.

[M ANSARI]

The way to prevent fortress positions is to use some sort of Monte Carlo analysis. You will quickly get a tree that will see that certain moves make no progress while others keep you chances alive. Once Monte Carlo can be a tool that can be used in real time then the last bastion for human chess will be gone.

[bob]

The way to prevent fortress positions is to use some sort of Monte Carlo analysis. You will quickly get a tree that will see that certain moves make no progress while others keep you chances alive. Once Monte Carlo can be a tool that can be used in real time then the last bastion for human chess will be gone.

Unfortunately, when that happens it won't be needed, because then programs will be able to search deep enough to see the end of the game anyway. In any position where you can not search deeply enough to see the end, you will get wrong results, whether it be from a monte-carlo simulation or a real search.

[Uri Blass]

The way to prevent fortress positions is to use some sort of Monte Carlo analysis. You will quickly get a tree that will see that certain moves make no progress while others keep you chances alive. Once Monte Carlo can be a tool that can be used in real time then the last bastion for human chess will be gone.

monte carlo may give you misleading results.

You may fail to win some position only because you do not search deep enough in the monte Carlo search.

[Edmund]

The way to prevent fortress positions is to use some sort of Monte Carlo analysis. You will quickly get a tree that will see that certain moves make no progress while others keep you chances alive. Once Monte Carlo can be a tool that can be used in real time then the last bastion for human chess will be gone.

I don't think Monte Carlo analysis is a solution to this problem. Keeping a fortress stable means for the defending side to follow certain rules, for example don't leave a certain region with the king. How would MC spot these rules?

[bob]

The way to prevent fortress positions is to use some sort of Monte Carlo analysis. You will quickly get a tree that will see that certain moves make no progress while others keep you chances alive. Once Monte Carlo can be a tool that can be used in real time then the last bastion for human chess will be gone.

I don't think Monte Carlo analysis is a solution to this problem. Keeping a fortress stable means for the defending side to follow certain rules, for example don't leave a certain region with the king. How would MC spot these rules?

You hope you search and produce enough games that you find good moves purely by accident and then discover they win. But if this is a depth issue, you have to play fast games to play enough to use for the MC method. And that means even shallower searches than normal, such that you might _never_ realize that you have to make this move now, so that 25 moves in the future your opponent can't escape via a fortress.

[M ANSARI]

The "real time" MC I envision is something that is running at incredibly fast speeds ... maybe even by a specialized node set or a specialized Quad SLI VGA card. Something that could simulate thousands or millions or hundred of millions of games very quickly. MC works very well and using my Octa I was able to find moves within a couple of hours of MC analysis that would never have been found by any engine. MC will quickly spot areas of no progress or will find very difficult to find drawing sequences or find obscure hard to find wins. Real time MC analysis would be used only as a node to give the main engine ideas of what lines to search and not as an analysis all by itself. It should work alongside the already very powerful main hardware.

[bob]

The "real time" MC I envision is something that is running at incredibly fast speeds ... maybe even by a specialized node set or a specialized Quad SLI VGA card. Something that could simulate thousands or millions of games very quickly. MC works very well and using my Octa I was able to find moves within a couple of hours of MC analysis that would never have been found by any engine. MC will quickly spot areas of no progress or will find very difficult to find drawing sequences or find obscure hard to find wins. Real time MC analysis would be used only as a node to give the main engine ideas of what lines to search and not as an analysis all by itself. It should work alongside the already very powerful main hardware.

I don't see how a ton of very shallow searches is going to give any guidance if the problem is such as those produced by a fortress position. Once it is set up, it is impossible to break down. And it often takes forever to realize this has happened.

The problem with fortress positions is that we can't search deep enough to avoid them because we need to recognize we have reached a no-progress type position first. Shallower searches are not going to help here at all.