| |
BGonline.org Forums
A few comments
Posted By: Timothy Chow In Response To: Very impressive result from DeepMind team (AP)
Date: Wednesday, 13 December 2017, at 6:25 p.m.
I'm grateful to Michael, who mentioned on rec.games.backgammon that this topic was being discussed here on BGO. I have a few comments to make.
First of all, I think that the best chance of getting the DeepMind team to take on backgammon is to explain to them that current BG bots suck at snakes (and related issues such as "wild" containment positions). Researchers are attracted to unsolved problems like ants to sugar. They're obviously not interested in making a few bucks at the local chouette or winning the nonexistent Backgammon Bot World Championship. What they want is the cachet of being able to say that their algorithm easily solves a problem that has been open for decades. Of course, even with this incentive, they're probably not interested in backgammon because public interest in backgammon is low. The best chance IMO is to get them interested in poker and have them tack on backgammon to poker as an afterthought. Even this ploy might not work, but in the worst case, I think that eventually the methodology will make its way into the public domain and anyone who has the time and inclination will be able to create AlphaGammon themselves (as with all technologies, the computational power needed for the training process will decline over time, and a few bucks of cloud computing power will eventually suffice).
As for whether to be impressed by the achievement, I think it is a mistake to focus on the ways in which they crippled Stockfish in order to get a more sellable headline. Yes, that was a cheap shot, but you shouldn't be falling for clickbait anyway. The point is that the novel combination of Monte Carlo tree search and convolutional neural nets has now been demonstrated to have phenomenal power. Neural nets have been tried on chess before, without success. Variations of Monte Carlo tree search have been tried as well, again without being able to beat alpha-beta search. The fact that such an approach is able to handle highly tactical issues such as joseki, ko fights, and chess combinations in addition to overall positional judgment—better than or at least roughly comparable to existing programs—is nothing short of miraculous. Regardless of whether the Stockfish sides of the published games are reproducible, the AlphaZero sides of the games amply demonstrate that AlphaZero is breaking new ground. It's irrelevant whether the current version is actually better than Houdini or Komodo or Stockfish with all the bells and whistles attached. The key point is that this is a new idea and good new ideas are hard to come by.
For this reason, I have no doubt that AlphaGammon will be able to handle snakes easily. The Monte Carlo tree search approach means that it should stumble into the snakey region of the game space more readily than conventional backgammon neural net training methods will, and I see no reason to doubt that, once in that region of the game space, it will figure out what's going on. As mentioned above, the ability of AlphaZero to handle highly tactical issues means (I predict) that rolling primes around the board should be a snap for it. From a backgammon player's perspective, I find this exciting because it would mean that the one known gaping hole in the current bots would be largely plugged, and we could gain increased confidence in superbackgame evaluations, and by extension, less extreme and more practically relevant backgame and proto-backgame evaluations.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.