🚀 go-pugleaf

The weighting of boxes method you describe might give a reasonable opponent.
I used a very similar method for the game of Qubic (like 3-D tic-tac-toe, but
get 4 in a row). The resulting algorithm was extremely fast (no time delay)
and in a software engineering class here at school, it finished in 3rd place
out of about 30 programs. Most of the competing programs used alpha-beta
pruning with look ahead. The winning program used 9 level lookahead with
alpha-beta pruning. I was surprised the no-lookahead method did so well...

Another interesting method would be to implement a neural net that could
learn. A paper was recently published about using Temporal Difference algs
for learning to play backgammon, with extremely good results. (Unofficially
beat world class level players more than 50% of the time.)

For a description of TD, see "Machine Learning", volume 3 (I think), from
1988. Its seems to be a very good neural net learning method...


-Brian