quarter ply , beyond its normal stength

[mainsworthy]

when an engines has reached its full ply search consider a very different strategy with it, it uses two algos, randomly choosing between them(percentage)! this gives random play without sacrificeing strength, and improves an engine to reach beyond its ply, by a quater ply

does this make sence to you?

any way Tournament Mainsworthy 26 has this feature check Mainsworthy on Leos site

http://wbec-ridderkerk.nl/index.html

[mainsworthy]

sorry but I need to clarify.

when an algo is to be executed ramdomly decide say a 70% chance of executing a 6ply or 30% 7ply choice.

[mainsworthy]

or different algo types

[Dirt]

or different algo types

Wouldn't that mess up your transition table? I think the move order found for a shallower prior search would be less likely to be correct in a deeper search using a new algorithm.

[bob]

or different algo types

Wouldn't that mess up your transition table? I think the move order found for a shallower prior search would be less likely to be correct in a deeper search using a new algorithm.

I think the idea is basically flawed, because trans/ref stuff depends on predictable depth as you mentioned. But randomly doing a deeper search seems to make little sense to me in that most of the tree branches are irrelevant anyway and extending those seems pointless.

[mainsworthy]

Yep Greg, you probably have to be carefull to fit it together in some circumstances! the idea is to give random play with extra playing strength, so two chess programs with there own tables etc.. need to work side by side, randomly switching (maybe 75 percent, for a normal move, and 25 percent chance of hitting a powerfull move with more thinking time)

your thinking about current practice of how tables etc.. are stored, this idea may need a whole new approach, to squeeze out a quarter ply.

this idea may suit some engines, and maybe not so good for others, but you can with corrected code get a quarter ply, eg: if your engine gets 2000 elo rateing, but has extra time left to make that elo, but not enough time to perform a full width search, you may be able to give it a partial new and maybe different algo and eval search and get a few more elo!

I dont know if im correct, for your thinking, or indeed any other coders, but my engine plays better(even though I didnt have enough time on the clock for a full width search)

anyway greg thanks for that problem pointed out.

[michiguel]

or different algo types

Wouldn't that mess up your transition table? I think the move order found for a shallower prior search would be less likely to be correct in a deeper search using a new algorithm.

I think the idea is basically flawed, because trans/ref stuff depends on predictable depth as you mentioned. But randomly doing a deeper search seems to make little sense to me in that most of the tree branches are irrelevant anyway and extending those seems pointless.

Unless it is pseudo-random and it behaves exactly the same every time it reaches the same position. For instance, the number to make the decision could be picked from the zobrist key of the position.

I am not saying the idea is good or not, but I am saying that it is possible to make sure that the hashtable is not messed up.

The idea may need to be explored. I do not see why discarding an unused fraction of a ply should be a better procedure than using it to increase the chances to go deeper.

Miguel

[mainsworthy]

or different algo types

Wouldn't that mess up your transition table? I think the move order found for a shallower prior search would be less likely to be correct in a deeper search using a new algorithm.

I think the idea is basically flawed, because trans/ref stuff depends on predictable depth as you mentioned. But randomly doing a deeper search seems to make little sense to me in that most of the tree branches are irrelevant anyway and extending those seems pointless.

hi robert, i wasnt thinking of usual programing principles, i was thinking redesigning tables etc..

if you dont think it would work, you may be wiser than me but i think it works for me so far, but I only my weak engine to play with, you probably got more experience.

[bob]

or different algo types

Wouldn't that mess up your transition table? I think the move order found for a shallower prior search would be less likely to be correct in a deeper search using a new algorithm.

I think the idea is basically flawed, because trans/ref stuff depends on predictable depth as you mentioned. But randomly doing a deeper search seems to make little sense to me in that most of the tree branches are irrelevant anyway and extending those seems pointless.

Unless it is pseudo-random and it behaves exactly the same every time it reaches the same position. For instance, the number to make the decision could be picked from the zobrist key of the position.

I am not saying the idea is good or not, but I am saying that it is possible to make sure that the hashtable is not messed up.

The idea may need to be explored. I do not see why discarding an unused fraction of a ply should be a better procedure than using it to increase the chances to go deeper.

Miguel

I'd agree _if_ there is some basis for making the decision to go deeper here as opposed to over there. Doing it randomly is not something I would consider at all.

[mainsworthy]

why Im thinking about this quarter ply , isnt because of one search, its because of average time per 20-40 moves , one quarter are power searches and the rest normal searches. so if you did all power or deeper searches, they wouldnt be enogh time on the clock, but there is time to do some deeper searches.

robert I think random 75/25, because your engine will be more unprodictable, and interesting without weakening it, infact it strengthens it(maybe).

thankyou Miguel I can see you know your stuff.