In recent topic, there was discussion about training a bot with Snake.
We all know that bots don't know how to play Snake because it's too much different from "real backgammon". (Tom said that bot could probably be good at either backgammon or snake, not in both)
I'm wondering if something between Snake and "real backgammon" backgame would be good for training a bot to better play backgames?
For example:
 is Player 2score: 0 pip: 170 | ||||||||||||||||
 | Money session Beaver | |||||||||||||||
| 6 | pip: 284 score: 0  is Player 1 | |||||||||||||||
| XGID=-a---aB-A---fB---c-d--BBBD:0:0:-1:00:0:0:2:0:10 | ||||||||||||||||
| on roll, cube action? | ||||||||||||||||
eXtreme Gammon Version: 1.20
Modifications from Snake:
- Yellow - only four checkers on the bar, 2 checkers on both midpoint and 6pt, blot on 8pt
- Brown - four checkers on 6pt, splited back checkers, six checkers on midpoint
- pipcount difference is ~100 pips, not ~200 like in Snake
| Analyzed in 3 ply | |
| Player Winning Chances: | 60,01% (G: 38,05% B: 9,98%) |
| Opponent Winning Chances: | 39,99% (G: 1,96% B: 0,07%) |
| Cubeless Equities | |
| No Double: | +0,660 |
| Double: | +1,321 |
| Cubeful Equities | |
| No Double: | +0,915 (-0,085) |
| Double/Take: | +1,049 (+0,049) |
| Double/Drop: | +1,000 |
| Best Cube action: Double / Drop | |
eXtreme Gammon Version: 1.20
| Analyzed in 4 ply | |
| Player Winning Chances: | 62,69% (G: 40,65% B: 11,30%) |
| Opponent Winning Chances: | 37,31% (G: 4,00% B: 0,23%) |
| Cubeless Equities | |
| No Double: | +0,731 |
| Double: | +1,462 |
| Cubeful Equities | |
| No Double: | +0,972 (-0,028) |
| Double/Take: | +1,208 (+0,208) |
| Double/Drop: | +1,000 |
| Best Cube action: Double / Drop | |
eXtreme Gammon Version: 1.20
| Analyzed in 5 ply | |
| Player Winning Chances: | 59,91% (G: 39,47% B: 10,43%) |
| Opponent Winning Chances: | 40,09% (G: 2,68% B: 0,10%) |
| Cubeless Equities | |
| No Double: | +0,669 |
| Double: | +1,339 |
| Cubeful Equities | |
| No Double: | +0,942 (-0,058) |
| Double/Take: | +1,066 (+0,066) |
| Double/Drop: | +1,000 |
| Best Cube action: Double / Drop | |
eXtreme Gammon Version: 1.20
| Analyzed in 6 ply | |
| Player Winning Chances: | 61,74% (G: 41,57% B: 12,60%) |
| Opponent Winning Chances: | 38,26% (G: 3,96% B: 0,22%) |
| Cubeless Equities | |
| No Double: | +0,734 |
| Double: | +1,469 |
| Cubeful Equities | |
| No Double: | +0,978 (-0,022) |
| Double/Take: | +1,209 (+0,209) |
| Double/Drop: | +1,000 |
| Best Cube action: Double / Drop | |
eXtreme Gammon Version: 1.20
| Analyzed in XG Roller | |
| Player Winning Chances: | 59,06% (G: 38,92% B: 10,39%) |
| Opponent Winning Chances: | 40,94% (G: 2,47% B: 0,09%) |
| Cubeless Equities | |
| No Double: | +0,649 |
| Double: | +1,278 |
| Cubeful Equities | |
| No Double: | +0,913 (-0,079) |
| Double/Take: | +0,992 |
| Double/Drop: | +1,000 (+0,008) |
| Best Cube action: Double / Take | |
eXtreme Gammon Version: 1.20
| Analyzed in XG Roller+ | |
| Player Winning Chances: | 60,44% (G: 40,07% B: 11,25%) |
| Opponent Winning Chances: | 39,56% (G: 4,09% B: 0,23%) |
| Cubeless Equities | |
| No Double: | +0,679 |
| Double: | +1,162 |
| Cubeful Equities | |
| No Double: | +0,906 |
| Double/Take: | +0,859 (-0,047) |
| Double/Drop: | +1,000 (+0,094) |
| Best Cube action: No Double / Take | |
eXtreme Gammon Version: 1.20
| Analyzed in Rollout | |
| No Double | |
| Player Winning Chances: | 60,86% (G: 40,09% B: 11,47%) |
| Opponent Winning Chances: | 39,14% (G: 2,80% B: 0,10%) |
| Double/Take | |
| Player Winning Chances: | 63,04% (G: 43,82% B: 15,57%) |
| Opponent Winning Chances: | 36,96% (G: 2,65% B: 0,11%) |
| Cubeless Equities | |
| No Double: | +0,704 |
| Double: | +1,654 |
| Cubeful Equities | |
| No Double: | +1,030 |
| Double/Take: | +1,291 (+0,261) |
| Double/Drop: | +1,000 (-0,030) |
| Best Cube action: Too Good to Double / Drop | |
| Rollout details | |
| 648 Games rolled with Variance Reduction. Dice Seed: 1296 Moves and cube decisions: 3 ply | |
| Confidence No Double: | ± 0,109 (+0,921...+1,138) |
| Confidence Double: | ± 0,199 (+1,092...+1,489) |
| Double Decision confidence: | 70,4% |
| Take Decision confidence: | 99,8% |
| Duration: 22 minutes 09 seconds | |
eXtreme Gammon Version: 1.20
So, we have here: no double, double/take, double/pass and too good to double.
Also, it's good to see how different plies play different moves for rolls like 61 (hit, 7 point, slot) for Brown.
Would training a bot with positions like this one help a bot to better understand "real backgammon" backgames?
I agree with Timothy, when he says: It's certainly true that one cannot hope to cover all possible positions. However, deep backgames, while rare, are not that rare. It's not like only one in a billion games ever turns into a deep backgame.