This article originally appeared in the August 2001 issue of GammOnLine.
Thank you to Kit Woolsey for his kind permission to reproduce it here.

When God is Wrong

By Kit Woolsey
When I play a match which has been recorded online, I run it through Snowie to find where my errors are. Snowie can zip through a match with full 3-ply analysis (not rollouts), huge search space, 100% speed, in 15 to 30 minutes. I then replay the match, and Snowie will tell me what my major errors were.

Most of the time I take Snowie's word when it says I have made an error. The program is damn good, and is usually correct. For some positions, however, I am not convinced that Snowie is right. I subject these positions to a rollout.

Because I want the answers quickly, the rollout parameters I use are not good enough so the results can be trusted with complete accuracy. My parameters are:

2-ply for play decisions
72 trials with variance reduction
Super-tiny search space
Cubeless
No truncation until bearoff data base is reached

With these parameters, I can usually get rollout results in a few minutes.

The 2-ply rollout is a big improvement over 1-ply. Snowie plays the checkers considerably better, so the results are more trustworthy. 72 trials with variance reduction usually comes out to the equivalent of about 1000 or 1500 games. While I don't claim to understand the variance reduction technique fully, I have been impressed with the results -- they seem intuitively to be quite accurate. I have also been assured by several knowledgable people that the technique is sound. So, I have learned to trust variance reduction. My sample size isn't too great, but the 95% confidence level is usually smaller than .05 so the results don't figure to be too far off.

About my other parameters: Super-tiny search space is generally sufficient. Snowie will usually find the best play among its top three 1-ply choices, and even if it doesn't most of the time that won't have a serious effect on the rollouts. I run them cubeless because cubeful rollouts take more time and lend another element of uncertainty to the results if Snowie is misevaluating the general position. I run the rollouts to completion for the same reason -- a major Snowie misevaluation can lead to distorted results with truncation. The bearoff database is safe, so there is no harm truncating when that is reached.

Obviously more accuracy can be achieved with different parameters, such as 3-ply rollouts, more trials, and larger search space. There is a tradeoff in time for each of these improvements, and I find the parameters I use generally give decent results quickly.

Most of the time when I roll out my errors, the rollout echoes Snowie' 3-ply opinion and I have egg on my face, with the only new point of interest being just how large an error I made. On occasion, however, I am vindicated, as Snowie reverses its opinion on the rollout. It should be noted that there tends to be a bias in the rollouts in favor of Snowie's 3-ply opinion. The reason is that Snowie is basing its 3-ply evaluation partly on how it thinks the position should be handled, so in the rollout it will back up this decision and play the position that way right or wrong. Thus, when Snowie does reverse itself in the rollout, the position is noteworthy.

I have gathered a collection of such positions from matches I played during the last few weeks. There is nothing particularly unusual about most of these positions that I can see. They are your normal everyday sort of position which, for some reason, Snowie appears to have misevaluated. These are not small technical errors. They are big conceptual errors, where the difference between Snowie's 3-ply evaluation and the rollout results is quite large.

Why does Snowie make mistakes? For the same reasons that we humans do. Snowie analyzes a position by assessing and weighing various parameters, just like we do. It is usually better equipped to take all these parameters into account and weigh them properly than we mere mortals, which is why its judgment is usually better than ours. However, a position can pop up where, for some reason, Snowie's weights are not an accurate representation of what is important. When this happens, Snowie can make a big error in exactly the same way that we do.

What is interesting is that in many of these positions, my choice of play was based on principles which I have learned from seeing Snowie's opinion of similar positions. Thus, finding Snowie getting them wrong is often quite surprising.

The positions are not in any particular order -- just as I found them.

164








130

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 0

A simple concept. Is it worth breaking the five-prime in order to make the ace point? Making the ace point will certainly win more gammons. If White stays on the bar, I will have time to pick up White's other blot and bring my men around with only the two-point game to be feared. However, if White enters quickly, things may not go well. My bar point will be lost, and I probably won't be able to recover it. White will be able to escape with fives and sixes, and he has enough outfield control so he can keep his position in one piece. I have a lot of checkers to bring around, and one of them might not make it home. It is easy to see how White can win this game frontwards.

Holding the five-prime with 24/20, 15/13 looked to me to be better. It would not be a disaster if White hit the blot on my ace point, since White would still have a lot of work to do to extricate everybody. In addition, I would immediately be attacking White's blot on my eight point. But was my thinking too much 1970's prime oriented? Is the brute force approach better here? Snowie thought it was.

3-ply evaluation

7/1             .860
20/16, 15/13    .801
The rollout told another story.

Rollout results

24/20, 15/13    .752
7/1             .722
Why did Snowie get this wrong? Sometimes Snowie seems to underestimate the value of primes and timing problems. Apparently Snowie didn't see how likely White's position is to crunch if I just hold the five-prime and sit on the position.

164








124

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 0

This is the same game, one roll later. The theme is exactly the same. Once again I have the choice between making my ace point or holding the bar point.

3-ply evaluation

20/18, 7/1      .990
20/12           .881
This time Snowie not only said that making the ace point was correct, it assessed my pure play as a big blunder. Once again I was vindicated by the rollout.

Rollout results

20/12           .928
20/18, 7/1      .888
I have often been dinged for making the same conceptual error on several consecutive plays. It is nice to see Snowie falling into that same trap for a change.

92








109

0123456bar789101112

0123456bar789101112
White 2



7 point match




Blue 2

This one looked simple enough. I can't keep my full prime, so I might as well cover the two point with 9/2 and try to win from there. I may be able to scramble my back checker up and out if White stays on the bar, and if White enters I still have my five-prime so I have a reasonable chance to win the priming battle. Could the super-pure 8/3, 5/3 be right? Nah. Getting hit would be a major disaster, since White would be ready to escape with a six while I am stuck on the bar. If White doesn't hit I still have to cover the blot on the two point, and even then I will have work to do. Making the three point is a play I might have made several years ago, but the bots have taught me that super-purity isn't necessarily right in this sort of position.

3-ply evaluation

8/3, 5/3        .473
9/2             .264
What is this? Snowie not only likes making the three point better, but it dinged my play as a double whopper. The bot and the human have swapped places, with the bot making the super-pure play and the human the more tactical play. Could Snowie possibly be right on this one? I found it hard to believe.

Rollout results

9/2             .197
8/3, 5/3        .139
Much better! Why did Snowie get this so wrong, when all our experience with the bots tells us that we would expect it to lean towards 9/2. I really don't understand what happened. Yes there are some duplication factors after making the three point since I would need aces and maybe sixes on both sides of the board if White misses the shot, but Snowie is usually very good at picking up on relevant duplication. It looks like Snowie is simply overevaluating the importance of purity in this particular prime vs. prime battle. While playing pure is often correct, this simply isn't the time for it.

166








163

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 3

White on roll: Cube action?

I used to take too many cubes like this. Now I respect the power of a potential blitz when I am on the bar and am staring an attack in the face. In my favor is an even race, White having three men back, and White not having too much attack material up front. Still, White has enough power to do plenty of damage, and I figure to come out with an inferior position even if things go well. Also the match score made me more hesitant to take, since my recube vig would be less than normal due to my lead in the match.

3-ply evaluation

Cubeless equity: .453

Double, take:   .824
No double:      .673
Double, pass:  1.000
Ding, says Snowie. Easy take. Pass is a huge blunder.

Was this really the case? Should I go back to taking cubes such as this?

Rollout results

Cubeless equity: .554

Double, pass:  1.000
No double:      .771
Double, take:  1.049
Not so, says the rollout. White's position really is that strong. I would have a money take, but at the match score I appear to have a close pass. It is not clear to me why Snowie misevaluated this position so badly. The bots have consistently been punishing me for taking this type of position when I am on the bar and in danger of being attacked without much on my side of the board. I guess Snowie was overly influenced by the race and White's three men back, and somehow didn't see just how strong White's potential is. The key is that White doesn't have to play a blitzing game. He can play positionally if the dice so dictate, and he will still have a clear advantage. This apparently is what Snowie failed to see in its evaluation but what became clear in the rollout.

152








109

0123456bar789101112

0123456bar789101112
White 1



19 point match




Blue 4

Gotta fight for that five point, right? Don't worry about blots. 8/5*, 7/1 must be better than the stagnant 7/1, 4/1. If I let White anchor, my advantage goes down the drain. The gains from hitting are huge when it works, and if I am hit back it is no disaster. Or am I going back to 1970's thinking again?

3-ply evaluation

7/1, 4/1        .612
8/5*, 7/1.      .548
1970's thinking, says Snowie. You gotta walk before you can run. Make points first, and don't leave a million blots. I was almost convinced. Still, with 13 checkers in the attack zone, could the blitzing play really be that bad? I have often been dinged by Snowie for being too conservative in positions such as this. Had to check it out.

Rollout results

8/5*, 7/1       .602
7/1, 4/1        .531
A complete turnaround. The rollout indicates that the double-blot play is clearly superior. Apparently Snowie underestimated just how much ground I lose if White makes the anchor, and overestimated the cost of being hit back. I think the key is that I am so far ahead in the race that even if I am hit I will still have the lead. However, if White gets the anchor then White's position will play very comfortably. Snowie realizes all this, but just underestimated the extents involved.

148








173

0123456bar789101112

0123456bar789101112
White 13



19 point match




Blue 9

In the old days, I might well have played the spread-them-out play of 13/11*, 11/9, 6/4(2)*. After all, White has two checkers on the bar, so what bad can happen? The nine point is more valuable than the 11 point, and I am almost sure to get it if I slot it.

Not any more. The bots have tought me that assets are everything. In particular, when I am short on builders it is important to grab any key point that I can. A bird in the hand is worth two in the bush. Those bad things which can't happen just might happen. I want that 11 point nailed down for sure.

3-ply evaluation

13/11*, 11/9*, 6/4(2)*  .917
13/11(2)*, 6/4(2)*      .877
Not so, says Snowie. Back to the 1970's. Spread them out. Why?

Rollout results

13/11(2)*, 6/4(2)*      .940
13/11*, 11/9, 6/4(2)*   .878
Well, that's a lot more Snowie-like. The 11 point is very valuable with this structure of the four point made and the five point open -- almost as valuable as the nine point. That bird in the hand is worth two in the bush. Snowie knows all this. I am at a total loss as to why it preferred the alternative play.

136








150

0123456bar789101112

0123456bar789101112
White 15



19 point match




Blue 14

I am behind in the race, and White is threatening to escape his last back checker. In addition, White has a few threats on his side of the board. It seems natural to hit loose with 16/11, 6/4*, which conveniently unstacks my heavy six point as well as knocking White back. It is true that White has the slightly stronger board, but the other relevant bold play vs. safe play critera appear to argue for the bold play.

3-ply evaluation

16/11, 13/11    -.510
16/11, 6/4*     -.561
Not so, says Snowie. Just lock up that 11 point for now, play with one blot, and see what happens. Is this right?

Rollout results

16/11, 6/4*     -.526
16/11, 13/11    -.549
The rollout tells another story, placing the loose hit a bit ahead of making the 11 point. It is interesting that Snowie, who usually follows the bold play vs. safe play criterion religiously, went against its own principles for this position.

109








103

0123456bar789101112

0123456bar789101112
White 1



7 point match




Blue 4

Ten years ago I would have played 7/4 without giving it a second thought. Keep the strongest board. What other consideration could there possibly be?

The bots have tought me otherwise. Builders and flexibility are what holding games are about. In particular, a spare on the six point is golden when the opponent has your five point. Going behind the enemy anchor is death. There is almost always a clever switching play available which will allow you to keep your builders in front of the anchor without compromising the position too much. For this position, the clever play is 7/6, 4/2. That type of play has become so second-nature to me now that I made the play without giving it any thought.

3-ply evaluation

7/4             -.374
7/6, 4/2        -.420
Not this time, says Snowie. Holding the four point rather than the two point has priority here. I was surprised. Is there something different about this position?

Rollout results

7/6, 4/2        -.387
7/4             -.409
The rollout vindicated me. The results were close, which was no surprise since there can't be too much equity difference between the plays in any event, but the clever switching play which leaves the builder on the six point came out ahead. What puzzles me is why Snowie chose this position to ignore this theme. Perhaps the fact that I have a single blot on the other side of the board which is likely to be attacked convinced Snowie that it was more important to hold the four point in case I am hit and I hit back. There is some logic to this, but the thesis didn't stand up to the test.

139








82

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 1

White on roll. Cube action?

Obviously White has a powerful redouble, since if he can stay on the bar a couple of rolls Blue's position is likely to crack. I put a fair amount of effort into analyzing this problem, since we are talking about a recube and I am a bit ahead in the match. It is easy to see the bad scenario -- White flunks, I have to play, White flunks, and my board becomes history. I am not at the edge of the prime, so even if White is forced to break his prime I still have a long way to go. Do I have sufficient winning chances to justify a take?

3-ply evaluation

Cubeless equity: .475

Redouble, pass: 1.000
No redouble:     .828
Redouble, take: 1.018
Clearly a money take, but at the match score Snowie evaluates it as a close pass. I wasn't convinced. The key is that if White enters at all and can't escape immediately, it is good-bye to the six-prime. Once the six-prime breaks he won't be able to put it together, and I will have a good chance to scramble out. Even if my board cracks a bit and I lose my six point, I will still have plenty of play with a solid four-prime and a big racing lead. It will be White who has to prove himself. And as long as I have my six point, most of White's entrances force him to crack. Was all this wishful thinking?

Rollout results

Cubeless equity: .418

Redouble, take:  .931
No Redouble:     .797
Redouble, pass: 1.000
It is still close, but the rollout dropped White's cubeless equity by over .05 and that was enough to give me a take. I wouldn't call this a big error by Snowie, since its evaluation was pretty close to the rollout results. What interested me about the position was that by applying logic, calculations, and estimations to the position I was able to come up with the correct conclusion, while Snowie with its general evaluation process failed. Until another giant programming step is made this is the area where we humans still have the edge over the neural nets, so we better make the most of it when the right positions arise.

143








162

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 3

When ahead in the race, race. When behind in the race, don't race. Time and time again the bots have pounded this theme home. In this position, I am behind in the race. 24/21 is a racing play. The blot on the 11 point isn't in much danger, but advancing the back checker allows White to play behind me safely while waiting to roll his eventual doubles which will clear the anchor and win the game. Staying back makes it difficult for White to find a safe play. In addition, 13/11, 6/5 both makes the possibly valuable 11 point and unstacks the heavy six point profitably. White has no inner board, so I'm not in much danger yet of being blitzed. The advanced anchor isn't that vital. Surely there can be no question about this one, right?

3-ply evaluation

24/21           -.448
13/11, 6/5      -.480
Wrong, says Snowie. I should dig in with the anchor and play from there. I was quite surprised. I was sure I had seen similar positions (after all this is only the third roll of the game), and the best defense against the early 6-6 was usually to hang back and wait. Did I have this concept wrong?

Rollout results

13/11, 6/5.     -.449
24/21           -.480
The rollout backed me up. The swing isn't huge, but it is significant. What is unusual is that there was such a big difference between the evaluation and the rollout for two different plays this early in the game. My experience has shown that Snowie's 3-ply evaluation and rollout results are almost always in agreement in the very early stages of the game.

151








149

0123456bar789101112

0123456bar789101112
White 4



7 point match




Blue 3

The clock is ticking in this nearly symmetrical position. 24/23, 13/8 has the advantage of splitting the back men so as to give them more flexibility toward getting to the edge of the prime and escaping. 13/8, 6/5 has the advantage of putting an important spare on the five point for future board-building. Which is more important, offense or defense?

13/8, 6/5 follows the important principle of putting my checkers where I want them. I know that fifth checker on the six point belongs on the five point, while the split to the enemy two point may not turn out to be valuable. Several years ago I definitely would have played 13/8, 6/5. Hours of being beaten over the head by the bots have shown me that these positions need to be treated more carefully. It can be dangerous to have two checkers locked on the enemy ace point against a growing prime. I finally decided that the key factor was timing. I was ahead in the race, and after this play White would have three checkers on the midpoint to my two checkers. This indicated that a priming battle would be likely to favor White, so I should do something about that. Therefore, I opted for the splitting play, 24/23, 13/8.

3-ply evaluation

13/8, 6/5       -.042
24/23, 13/8     -.072
Wrong again, says Snowie. Offense before defense. Put your checkers where they belong. Will I ever learn?

Rollout results

24/23, 13/8     -.074
13/8, 6/5       -.175
The rollout vindicated me big time. Not splitting is quite a blunder. Snowie fully recognizes the importance of the split, but apparently was unable to grasp the timing considerations here. Moving the builder to the five point was too tempting. We have seen that one of the weakest conceptual part of Snowie's game is understanding timing considerations. Unfortunately, it is often the weakest part of a human's game also.

140








140

0123456bar789101112

0123456bar789101112
White 4



7 point match




Blue 3

This is the same game a couple of moves later, after both sides have made some improvements. Now I must choose between running with 22/13 and bringing builders down with 13/10, 13/7. In the old days, there was no way I would break that anchor. Much too dangerous. Better to bring down some badly needed ammo to build my own board. I'll roll that six when I need it.

The bots have told me time and time again how wrong this is. Sixes just don't grow on trees. Get that checker out while you can. Don't be afraid to leave the anchor -- the bad guy can't both prime you and attack you at the same time. And so on, and so on. My training had tought me to play 22/13.

As I was reaching for the checker on the 22 point, I had second thoughts. Fours and fives were White's worst numbers -- why should I make them good for White? I had spares to play with and attacking possibilities, so I wouldn't have to spring a back checker so quickly. It is White who has the stronger inner board -- I don't want to get battered around. My attack force is awkward -- I need builders to improve my inner board. After all this thinking, I finally changed my mind and played 13/10, 13/7.

3-ply evaluation

22/13           -.079
13/10, 13/7     -.194
Yep, another blunder. When will I ever learn? It seemed almost pointless to roll this one out, but just to kick myself in the head again ...

Rollout results

13/10, 13/7     -.089
22/13           -.092
Wait a minute! Not so clear. The rollout has the two plays as about equal. Maybe there was a method to my thinking (madness?) after all. I can understand Snowie liking the thematic run. Apparently it failed to weigh in the importance of offensive builders here. If I can build another inner board point before White escapes, I may not have to worry about a priming battle.

117








102

0123456bar789101112

0123456bar789101112
White 0



7 point match




Blue 1

This one seemed easy enough. I know I want my three point, so let's make it. Sure I am leaving a few more shot numbers, but at least many of these hitters come into range of my checkers on White's bar point so I will have return shots. Sure I leave more hit and cover numbers, but even that I can live with. Also, staying on the midpoint allows me to potentially link there if I roll a five next turn, while 13/8 splits my outfield resources into two. Communicate! Seems like a no-brainer.

3-ply evaluation

13/8            +.042
6/3, 5/3        -.058
A blunder? Now that is hard to believe. Why? That didn't make much sense to me. I anxiously awaited the rollout results on this one.

Rollout results

6/3, 5/3        -.006
13/8            -.047
Now that's more like it! Human logic and intuition clearly won out here. Frankly, I can't see any reason why Snowie downgraded making the three point so badly. Perhaps the loss of spares on the five and six points. Perhaps Snowie was overly impressed by the few extra hit and cover numbers. Perhaps Snowie failed to recognize the importance of potentially being able to make the midpoint and bring the checkers around together. This isn't like Snowie, as I have seen in the past it understands these concepts quite well. Something just went wrong here. I wonder if the fact that White didn't own his six point somehow caused Snowie's evaluator to go haywire, since most of the contact positions in Snowie's training involve both sides owning their six point.

116








139

0123456bar789101112

0123456bar789101112
White 4



7 point match




Blue 6

Note the match score -- that may be relevant here. I can't use a gammon, but my opponent can. I have the choice between leaving my anchor with 20/14(2), 7/1(2) or holding the anchor and playing the cramped 14/2, 7/1(2). The race is now close. My racing chances aren't great holding the anchor, but I may be able to scramble the back man out and over in time for a race. If I leave the anchor, the back checker is stranded -- with more gammon danger. White could easily build a full prime and walk it home. Holding the anchor seemed clear.

3-ply evaluation

20/14(2), 7/1(2)   -.388
14/2, 7/1(2)       -.451
Snowie dinged me rather harshly for not getting out when the getting is good. I found that a bid hard to swallow. Sure my play leaves my position somewhat cramped, but to just leave the back checker stranded? Granted White will have to use his next roll to get his back checker safe, but after that he will be free to roll the prime home.

Rollout results

20/14(2), 7/1(2)   -.488
14/2, 7/1(2)       -.488
Well at least that makes more sense -- the plays came out even. I still think that holding the anchor has to be better, but the idea that leaving the anchor was way better was hard to believe.

91








92

0123456bar789101112

0123456bar789101112
White 5



7 point match




Blue 3

Blue on roll. Cube action?

At this match score, White can take very aggressively in non-gammonish positions. From White's point of view:

White passes: Ahead 5-4 (2 away, 3 away), 60% equity.
White takes and wins: Wins match, 100% equity
White takes and loses: 5-5, 50% equity.

White is risking 10% equity to gain 40% equity, so he is getting 4 to 1 odds on his take if there are no gammons. This position isn't quite gammon free, but there is very little gammon risk. White certainly has enough winning chances to take, with the combination of winning the race and hitting a shot. The real question is whether I should turn the cube. It a matter of market loss. If I make my four point, will I lose my market? If yes, I should double. If no, I should wait. I thought yes.

3-ply evaluation

Cubeless equity: .464

No double:      .722
Double, take    .697
Double, pass   1.000
Snowie didn't think it was a double, which obviously means that Snowie didn't think it was a market loss if I make the four point. I wasn't convinced. White is going to have a tough time winning if I make my four point.

Rollout results

Cubeless equity: .540

Double, take:   .897
No double:      .865
Double, pass   1.000
The rollout said otherwise. Note the big change in cubeless equity -- it went from .464 to .540. That was enough to convince Snowie that I had the market losers. I have no idea why Snowie evaluated White's position so highly. Yes three-point games always have a lot of play, but this seems to be carrying the theme to an extreme.

108








107

0123456bar789101112

0123456bar789101112
White 0



11 point match




Blue 0

The age-old problem of whether or not to break an anchor. It certainly must be right to make the three point in all variations, so I will play either 11/7, 6/3 or 22/18, 6/3. I used to be paranoid about breaking an anchor. Not any more. The bots have tought me otherwise. It is often better to make a run for it than to find yourself getting stuck on an anchor forever. What are the conditions here? I am ahead in the race. I have by far the stronger board. White has that blot on his ace point, so my blots will be almost immune if I play 22/18, 6/3. I am blocked on sixes where I am now, so if I don't move from the 22 point I may have a difficult time freeing the back men later. White's position is likely to improve, while mine can only get worse. How many more arguments do I need to break the anchor? I played 22/18, 6/3 with complete confidence.

3-ply evaluation

11/7, 6/3       -.499
22/18, 6/3      -.588
What is this? Breaking the anchor is a blunder? This went against everything I had learned. As you can imagine, I was very eager to see the rollout results on this one.

Rollout results

22/18, 6/3.     -.488
11/7, 6/3`      -.599
Yep, there was a blunder to be made. But the blunder wasn't breaking the anchor -- it was holding the anchor! This is about a big a turnaround as I have ever seen from the 3-ply evaluation to a rollout for a relatively normal position. I am at a complete loss as to where Snowie got its evaluation. Obviously it thought my winning chances holding the anchor are better than they actually are -- we have already seen that from the cube decision of the previous position which is a relatively similar position. However, it would take more than that to produce this kind of monstrous turnaround. It is just baffling.

132








103

0123456bar789101112

0123456bar789101112
White 1



11 point match




Blue 0

How ugly is ugly? In the old days, a play like 8/2, 6/3 would be laughed at. The bots have shown us that sometimes ugliness is beautiful, and that super-safe pay-me-later ugly plays are often correct. But can it be correct here? Stripping both the eight point and the six point, putting a checker on the three point out of play, and leaving a blot on the two point which is difficult to cover? No way! 13/4* keeps a perfectly balanced position, and at least has some real upside if I'm not hit back. White has a strong board, but his three point is still open so getting hit back isn't necessarily fatal. If I delay White may close that three point, and then getting hit will be completely fatal. It must be right to pay now, right?

3-ply evaluation

8/2, 6/3        .185
13/4*           .126
Not so, says Snowie. Ugly is beautiful. Well, it sure seemed ugly to me. Let's see how ugly a rollout thinks it is.

Rollout results

13/4*           .126
8/2, 6/3        .088
Another turnaround. It is interesting that the equity after 13/4* agreed exactly with the 3-ply estimate, but the equity for 8/2, 6/3 dropped almost .100. Ugly isn't so beautiful after all.

119








100

0123456bar789101112

0123456bar789101112
White 1



11 point match




Blue 0

In a continuation of the same game after some scrambling this pay me now or pay me later position arose. It sure looked right to me to pay now with 14/10, 13/9(3). I would be virtually gin if I survived. White's board, while strong now, would only be getting stronger in the next few rolls. If I paid later with 14/6, 13/5, my position would be completely stripped and I would be very likely to have to leave a shot in the near future anyway. I wouldn't have time to wait around for my eventual doubles. A classic example of when to pay now.

3-ply evaluation

14/6, 13/5      .246
14/10, 13/9(3)  .224
Snowie thought the plays fairly close, but definitely want for paying later. This had to be suspect. All the relevant factors pointed towards paying now.

Rollout results

14/10, 13/9(3)  .289
14/6, 13/5      .197
As I had thought. The rollout results indicate that paying later is a big blunder. This seemed quite clear to me at the time. Snowie is usually deadly accurate on pay-me-now or pay-me-later decisions, and this one looks like a run-of-the-mill example. I have no idea why Snowie got this position so wrong.

154








210

0123456bar789101112

0123456bar789101112
White 0



11 point match




Blue 3

White on roll. Cube action?

While Snowie plays backgames better than most people think, it doesn't evaluate them very well. I had no difficult taking in the above position; in fact, I doubt very much if White should have doubled. However:

3-ply evaluation

Cubeless equity: .601

Double, pass:  1.000
No double:      .929
Double, take:  1.096
Snowie evaluates this as a moderate pass. I don't believe a word of it. This is one position where I would be confident betting against Snowie's judgment. My position simply has too much play if handled correctly.

Rollout results

Cubeless equity: .505

Double, take:   .859
No double       .807
Double, pass   1.000
A drop of almost .100 in equity for White's side, making it an easy take. And this is only on a 2-ply rollout. This is the sort of position Snowie plays the backgame side considerably better with a 3-ply rollout. It wouldn't surprise me if the 3-ply rollout indicated that it wasn't even a double. However, if you were playing against Snowie 3-ply, you could have stolen a point.

153








195

0123456bar789101112

0123456bar789101112
White 3



11 point match




Blue 3

One thing which the bots have taught me is that even when playing a backgame or a likely backgame it isn't a good idea to have checkers needlessly sent back. You can't afford to run short of men up front, or you won't be able to build a strong board quickly. While the choice of plays here didn't figure to be too important, leaving unnecessary blots in my outer board with something like 20/14 didn't seem right. More accurate appeared to play 20/18, 7/3, at least starting the next inner board point in line rather than throwing the checkers to the wolves.

3-ply evaluation

20/14           -.495
20/18, 7/3      -.578
Much to my surprise, Snowie thinks 20/14 is far better than my choice. Perhaps it is the duplication of twos, or perhaps the extra outfield presence and the improved possibility of making the eight point or the bar point. I was rather surprised. 20/14 is the sort of play I used to make, but I have learned to avoid that approach. Now Snowie is telling me that it is far better than the more conservative approach. This is a pretty big statement by Snowie, particularly since this is the sort of position where it isn't likely to matter too much what you do because you will probably wind up with a good 1-5 structure or something like that whatever happens.

Rollout results

20/18, 7/3      -.478
20/14           -.528
Well, the rollout found my play to be superior. I'm not exactly sure whether that means too much in this type of position, but it certainly puts to question Snowie's claim that 20/14 is far better.

115








106

0123456bar789101112

0123456bar789101112
White 5



11 point match




Blue 4

In my youthful days I'm sure I would have unhesitatingly played 10/1*. Go for the throat! What blockade? I can leap that little thing with one easy bound. This is gammon city. What can possibly go wrong?

Now that I am older and (hopefully) wiser, having been dinged by the bots countless times for plays such as this, I know better. There is a lot which can go wrong. 4-4, 3-3, and 4-3 are legal rolls. Or White might hit back, and I might enter with some prime busting number -- you know when a little 2-1 or 1-1 busts your prime you have a lot of bad numbers floating around. 21/15, 10/7 is clear. It will be very hard for the roof to cave in, and I will still have plenty of opportunities to win the gammon.

3-ply evaluation

10/1*           .921
21/15, 10/7     .914
Snowie thought the plays close, but believed that the extra gammons from the loose hit were sufficient to justify the play. Even though I wasn't dinged for a big error, I thought this was worth rolling out because I was convinced that Snowie was way out to lunch.

Rollout results

21/15, 10/7     1.052
10/1*            .858
I was not disappointed in the rollout. Snowie's choice, which it would have made had it actually been playing, was almost a double whopper. We have seen that Snowie can have difficulties with timing considerations, but this one is pretty extreme. What is surprising is that Snowie is usually quite sensitive to the dangers of getting stuck and being forced to break the board. Here, the combination of the gammon potential and the false sense of security lulled Snowie into an awful play.

No doubt many of you readers have collections of similar Snowie blackouts. It might be a worthwhile idea to put them all together and see if we can form some idea about the types of positions where Snowie is likely to falter. In the meantime, remember that Snowie is not infallible. Even God can be wrong.

Return to: