|
BGonline.org Forums
The next major revolution in backgammon bots
Posted By: Michael Petch In Response To: The next major revolution in backgammon bots (Stick)
Date: Tuesday, 22 May 2012, at 5:12 a.m.
"In fact our bot has the best play slotting the 5 and 4pt with a 3-1...that's illegal but I'm sure you meant to type something else that made sense."
My followup post partially corrected that issue. I had to go to the rollout/trainting data. Opening 8/7 8/5 was best.
Bots are biased because of short term thinking (look ahead). What we actually have is a bias towards racing because we don't have the resources to do infinite rollouts with infinite ply look ahead during training. Are we absolutely sure that "perfectly" played backgames from start to finish are actually worse than a race? If you ask commercial bots (and most humans) the answer is generally "Yes" based on the information we have today. But this is based on incomplete knowledge, and biased neural nets.
*If* there is going to be a significant "evolution" in bots, it will likely be in the area of better understanding the "long game". That may be a new way of making AI's, speeding up bots, more computing resources, further lookaheads. Secondly, I agree with the posters that believe an evolution in bots would also be in them being to articulate to humans the "Why" of a move, not just what play is better.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.