Killer moves heuristic not working in my engine.

[adams161]

acutally the i want to be clear, the swap idea is
sort=1;
for(a=0; a<top; a++) // loop through moves
{
if(sort==1)
sort=order(moves, a, top)

rest of search
}


int order(int moves[], int bottom, int top)
{
/*

if i have moves 1 2 3 4 5 at moves[0] through moves[4]
I start looking for moves at bottom ( on first pass 0 on secon pass 1
and when i find it i swap so if best move is 4 next pass i have
4 2 3 1 5
but bottom places me at 2 for the first move i now look at on second pass
*/
int place=-1;
int score=-1000;
for(a=bottom; a<top; a++)
{

// now i go through whole list starting at bottom and score moves
if(moves[a]==capture)
if(capturesscore(moves[a])>score)
{
place=a;
score=capturesscore(moves[a])
// place always becomes the highest scoring move in moves from bottom to top
}
if(killerscore(moves[a])>score)
{
score=killerscore(moves[a]]);
place=a;
}


}// end for bottom to top for loop

if(score>-1000)// we found a move
{
swap(place, bottom);
return 1;
}

return 0; // once 0 is returned we no longer call order and assume list is fully sorted
}// end function order

thats not exactly how i do it and there may be more effecient ways to do it but its the idea.

it garantees two things, one a smooth transition from one move ordering scheme to the other
and two, no moves are played twice


Mike

[mathmoi]

one interesting stat to keep when debugging killer moves, is when you play a kilerr move have 2 counters, one is the number of times you play a killer move in search the other is the number of times this played killer move produces a beta cuttoff.

I.e if you might play 500 killer moves and get 100 beta cuttofs ( just making up the numbers ). if you see something funny, like your playing not X killer moves but some huge amount, thats a problem. if you are not getting any cuttoffs at all, thats a problem.

Mike

Hi Mike (and others),

I just tried to extract the killer move success rate of my engine like you suggested. I get a sucess rate (killers that caused cutoff / killer tried) of about 45%. This seems low to me, even if I don't know what to expect.

What is an acceptable success rate?

[adams161]

you're not ordering captures as killers are you? you indicated you played captures before killers so not sure what would happen in your engine if killers could also be captures.

Mike

[mathmoi]

Hi Mike,

No, killers can't be captures or promotions, I test for this before I add a move to the killer list.

I'm begining to think that I have a serious bug in my engine making move that were good at a positions no longer good at an other. Now, I'm thinking about a way to "dump" the search to a file so I can examine what's going on at every nodes.

Thanks for your help,

[jwes]

one interesting stat to keep when debugging killer moves, is when you play a kilerr move have 2 counters, one is the number of times you play a killer move in search the other is the number of times this played killer move produces a beta cuttoff.

I.e if you might play 500 killer moves and get 100 beta cuttofs ( just making up the numbers ). if you see something funny, like your playing not X killer moves but some huge amount, thats a problem. if you are not getting any cuttoffs at all, thats a problem.

Mike

Hi Mike (and others),

I just tried to extract the killer move success rate of my engine like you suggested. I get a sucess rate (killers that caused cutoff / killer tried) of about 45%. This seems low to me, even if I don't know what to expect.

What is an acceptable success rate?

Is killer tried counted for all nodes or only for nodes that fail high?

[mathmoi]

one interesting stat to keep when debugging killer moves, is when you play a kilerr move have 2 counters, one is the number of times you play a killer move in search the other is the number of times this played killer move produces a beta cuttoff.

I.e if you might play 500 killer moves and get 100 beta cuttofs ( just making up the numbers ). if you see something funny, like your playing not X killer moves but some huge amount, thats a problem. if you are not getting any cuttoffs at all, thats a problem.

Mike

Hi Mike (and others),

I just tried to extract the killer move success rate of my engine like you suggested. I get a sucess rate (killers that caused cutoff / killer tried) of about 45%. This seems low to me, even if I don't know what to expect.

What is an acceptable success rate?

Is killer tried counted for all nodes or only for nodes that fail high?

It's counted for all nodes. Are you telling me I should only count them for nodes that actually fail high in order to ignore All-nodes?

[Gerd Isenberg]

Hi Mike (and others),

I just tried to extract the killer move success rate of my engine like you suggested. I get a sucess rate (killers that caused cutoff / killer tried) of about 45%. This seems low to me, even if I don't know what to expect.

What is an acceptable success rate?

Is killer tried counted for all nodes or only for nodes that fail high?

It's counted for all nodes. Are you telling me I should only count them for nodes that actually fail high in order to ignore All-nodes?

I think 45% killer_cuts of killer_tried is quite fine, since one tries killers even at nodes, which turn out to be All-nodes (most often already expected), where no move fails high at all. If you pass expected node type as parameter of a zero window scout search of a PVS, you may also try

killer_cuts[node_type] / killer_tried[node_type], with node_type {all, cut}

for a different image ...