Corner gain, Othello

[Guest]

Archived from groups: comp.ai.games (More info?)

Hi,
My X & C functions "score" only when X,C Squares adjecent to an empty
corner.

I think It is the corner gain that is problematic.

When I add the cornerGain to the start of the game , The performance
seems to get worse.

I probably don't factor the components very good:
I multiply it by 20 just because It felt right, but I guess I need to
test it throughly to really see the impact.

I eventually omitted the CornerGain from all stages of the game since
I cound not find the right measurment for it.

Any Ideas?
Would you consider the program to be a moderate strength ?
(goes up to depth 9-12)

BTW , I forgot to mention I have a "CornerGrabber" that immidieatly
plays a move if the player has a corner to conquer.
Do you think it is a good behavior?


Thanks.

 

[Guest]

Archived from groups: comp.ai.games (More info?)

To find the best coefficients, you will need to run many competitions
between different configurations to find the winning one. This works
better when your evaluation function is good.

Considering X and C squares (weight squares strategy) helps a lot. Your
program will beat most beginners, and people who understand the concept
of X/C squares and mobility. It would probably lose heavily against
players who start learning a bit more about edge strategy.

Corner grabbing is not always good (look up Stoner traps). You don't
need a CornerGrabber function, as you would find that with X+C+corner
weightage your program will be grabbing corners most of the time if it
is working correctly. If it is not, that means you are not canceling
out the negative effect of flipping the opponent's X/C square disc.
When your opponent puts a disk on a C square, he gets -20 (for
example), which means you get +20. When you play the corner, you score
the corner, but you lose the +20 for having opponent's disk on the C
square, which probably should not be the case.

Besides potential mobility (or frontier) which Richard has mentioned,
stability, or stable disc count is pretty important too. Your next step
should be to implement an edge pattern table. Implementing a good edge
table made my mobility based evaluator very strong, probably second
only to full pattern based evaluations. An edge table covers stable
disc and helps you avoid Stoner traps.