[Bug-gnubg] more about cubefull on/off and wierd results using noise in rollouts

[bob koca]

  I rolled out the position in which each player has 6 on the ace point in a one point match.
pos id is PwAAgB8AAAAAAA and match id is cAkgAAAAAAAA
 
  To 6 places the exact win% for the player on roll is .787809.  The evaluation function  ( is that from the database?) gives .787793.  I initially thought the discrepancy was due to rounding off in the number of rolls off probability calculations of 30.556% for off in 2 and 69.444% for off in 3 but that gives .787807
 
   Since the plays are forced the same result should be obtained with any skill level of players in the rollout but this is not the case. Note that if cubeful is on the results are all over the place. Also if cubeful is not used but the bearoff database is not used. 
 
  For all of these rollouts I used variance reduction, quasi random dice, and mersenne with seed 1000. 
If the table below is not formatted nicely see the attached excel file. 
 
  Can anyone make sense of these results? 
 
Bob Koca

On roll	Not on roll	RO win%	RO win%	RO win%	RO win%	Ro win %
		cubful ON	cubefull ON	CF OFF	CF Off	CF Off
		No B.O db	exact DB	No B.O db	Exact db	One sided
expert	expert	0.787809	0.787809	0.787809	0.787793	0.787793
adv (det noise)	adv (det nooise	0.723544	0.723544	0.723544	0.787793	0.787793
int (det noise)	int(det noise)	0.723544	0.723544	0.723544	0.787793	0.787793
expert	adv (det nooise)	0.812211	0.812211	0.812211	0.787793	0.787793
adv (det noise)	expert	0.699141	0.699141	0.699141	0.787793	0.787793
int (det noise)	expert	0.699141	0.699141	0.699141	0.787793	0.787793
int (not deterministic	int(not deterministic	0.723544	0.723544	0.788758	0.787793	0.787793
adv (but not cubeful)	adv(but not CF)	0.723544	0.723544	0.723291	0.787793	0.787793

---

Messenger Café — open for fun 24/7. Hot games, cool activities served daily. Visit now.

gnustudy.xls
Description: Binary data