Details of Program: Steenvreter


Written by: van der Werf, Erik


First appeared in 2007 Computer Olympiad.

"My impression is that Steenvreeter is a MC-program because it sometimes played moves that I find typical of MC-programs, but that it uses some very good tactics and L&D reading in the playouts. It played very strong in all tactical situations as far as I could see in a quick overview. Or perhaps it is a hybrid of a strong tactics engine and MC. Just a guess." – Magnus Persson.

The above quotation triggered this conversation on KGS (evdw is Steenvreter's author):

evdw: Magnus's impression is correct :-)
evdw: Tactics are handled by the UCT tree
evdw: It can play the whole game in a few seconds (if necessary)
evdw: About tactics you meant in the playouts?
evdw: If so, then yes. It has things like (simple) ladder searches in there...
tmc: I was wondering what gave it its tactical strength.
evdw: mainly the tree part
tmc: I don't understand UCT enough to follow you. Do you mean the stored tree has some biases built in?
pasky: evdw: how many montecarlo playouts do you do per move?
evdw: No, the tree converges to the minimax solution, which solves all the tacticts that are within the search horizon
evdw: now it is down at 29k/s
pasky: do you do some pattern matching or so in the playouts?
pasky: in that case 29k/s looks very impressive
evdw: of course, no strong program does uniform random these days...