|
||
|
Title: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:06pm That was a fun game. Time-consuming, but fun. Thanks to everyone for playing! Not sure whether this forum allows attachments, so just uploading them elsewhere and linking here. Here are my in-game notes written as I played each move, edited and cleaned up a bunch for readability/grammar/clarity: http://icosahedral.net/downloads/cyborgsharpvmob/ingamenotes.txt Here are the analysis trees that were produced in the process of choosing moves that go alongside the in-game notes. http://icosahedral.net/downloads/cyborgsharpvmob/trees.txt For convenience, a link to the finished game: http://arimaa.com/arimaa/gameroom/comments.cgi?gid=507629 |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:11pm And here is some move-by-move post-game commentary and review here. It's written by me mostly without having read through the mob forum threads, so if a few things I say here are wrong, I wouldn't be too surprised. Also with each Silver move is the rough cyborg eval I had during the game. This was Sharp's eval fuzzily minimaxed over the analysis tree, adjusted and biased by my own feelings about the position taking into account the analysis, and smoothing over Sharp's unfortunate even/odd ply eval bouncing. 1g Sensible opening. :) 1s I wrote a little bit of code to scrape the games archive for games between strong players and generate a simple book for Sharp. I then picked a setup in the resulting book that looked interesting. It turns out to be one of browni's openings (although not the most common). Cyborg eval: ~0 2g Not much to say. 2s Sharp's favorite move, and the book move. Cyborg eval: ~0 3g Seems also like a good move. It's interesting to advance the M directly up to the e, but any pushing or flipping of the M simply loses tempo, so silver can't do anything about it. 3s Given the game 4g, one wonders if mg7s hh7s ra7s ra8s would be better. Sharp was of very little strategic help on this and subsequent moves. But it was able to evaluate the fights that resulted from racing - if one side went all out and the other side counterattacked rather than defending. This proved very helpful, since it indicated which side would have to "back down" first. Also on this and the next move or two I relied heavily on the heuristics: * "Make sure that Silver's ahead in tempo if both sides launch symmetric attacks" - this heuristic prunes all a lot of silver elephant side-stepping, pushing and pulling of the M or H, etc. * "Maintain maximum flexibility" - Given that I didn't know what plan Gold would choose yet, I tried to play only steps that we would want to play anyways and that couldn't be easily made regrettable by Gold. Cyborg eval: ~0 4g The immediate Ha6 is nice! 4s Used the same heuristics as on 3s. Explored many racing lines deeper. Cyborg eval: ~0 5g Rh2n Rh1n was unexpected, but probably that's just because I don't have experience with opening strategy. Once on the board, it was clearly a nice move - the moment the M is too far to prevent hh3, seal it up with RR instead. 5s Tactically, spent time playing with Sharp interactively to explore the various Gold elephant crossings, the M advance, the d flip. With the M crossing west, it became a lot easier to plan ahead. Cyborg eval: ~0 |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:12pm 6g The strategic situation further clarifies, with the Gold E committing to be in the east, or else lose tempo crossing west. Gold's move is again very good. 6s This was a difficult turn. Rh2n Rh1n plus the Gold E so close makes it hard to make anything work around f3. For a long time I found nothing good, and Sharp continued to be strategically useless, although tactically it helped, and it also agreed that everything I explored was bad (-0.1, -0.2, etc.). Not knowing about hh4n Rh3n Rh4w hh5s being a "standard" sort of move in this situation, I was lucky to discover it almost just before I was about to make one of the other moves, when idly pondering the position for a final time. Sharp quickly confirmed that it was reasonable, and despite never suggesting it itself, in fact liked it more than all the other moves once I deepened on it. Cyborg eval: ~0 7g Gold plays one of the expected camel threats (the other being Ef5s mg5w Ef4w mf5s). 7s Despite expecting the camel threat, we didn't actually cyborg-search it very thoroughly. Fortunately, it turned out okay. Spent a lot of time analyzing what the fourth step should be, and the consequences for the upcoming fight. Explored many variations out 3-4 ply, with the deepest lines going to 7 or even 9 ply! (Plus 3.25 more from Sharp, since sharp reaches s13 fairly readily in an interactive explore. I would say that cyborg searching in this manner is at least moderately sensitive to tactics reaching out 6-10 ply). With the lucky break of finding 6s, the game up to this point has developed pretty much as desired - it's tactical and involves unbalanced MH attacks. The more tactical and racy the better, because with Sharp, that's clearly where our relative advantage is. Starting here and over the next move or two, Sharp took the lead - suggesting good moves for the main line rather than just in the racing branches, giving high-quality evals that I agreed with, etc. And Sharp remained in the lead for the rest of the game. From here on out, every move was chosen by building a large analysis tree using Sharp, and choosing the branch that Sharp liked the most, with a slight bias towards moves that I liked when Sharp was indifferent or was itself unsure. And where "liked the most" was not exactly minimax value, but rather a "fuzzy minimax" that I did in my head that sometimes behaved more like MCTS-style averaging, with a bias if one side seemed to have "more options" than the other, etc. As I noted on 10s during the game, with Sharp taking the lead for suggesting moves and evaluating, my role became: * Choosing where to have Sharp search (what branches of the tree to deepen on, based on where more tactical clarity was needed and would be valuable). * Sanity checking Sharp's results. When Sharp gave a weird result, I almost never outright overrode Sharp or disregarded its result - virtually always the approach was to deepen appropriately (on the suspect move that should be bad, or on the suspected refutation, on the position in general if it was a miseval), and see if Sharp would change its mind or if it would refute me back. * Re-running sharp with banned moves and manually adding a few moves near the root, just to ensure we had loose coverage of all of the major strategic flavors of move. Cyborg eval: ~0 8g Nice move. Somehow we were relying on the possibility of a camel tuck with 8s hh4s mg4e Rh5w mh4n and didn't see this Gold move that neatly prevented it. But Silver had other options too, obviously. 8s This was suggested by Sharp, and I found it to be a beautiful move. Pushing Rh3w instead of Rh3s is counterintuitive, but it works really well! It tangles up Gold from making threats in f3 and makes it very awkward for the H to reclaim g3. In retrospect, I think this move is the first hint that Silver might be edging into the lead (or turning the tables, if you think Gold was winning before). Although during the game I still thought it was even or maybe good for Gold. Cyborg eval: ~0 9g Reasonable. No idea if Gold had better, would need to do a lot more analysis from Gold's side. 9s The natural continuation, and Sharp likes it too. Maybe 9g was a slight inaccuracy. Cyborg_sharp started to like the position here - there was a somewhat thin set of lines where Gold maintained equality in the search, and deviations by Gold led to Silver gaining an edge. Starting on this move the deeper of the various analysis lines started to anticipate the m crossing to the west to take the Ha6 hostage, among other possibilities. In those variations Gold has regrets about 8g Ra4n, because once the m crosses, Ra5 makes it extremely hard to extract Ha6 cleanly. Cyborg eval: +0.1 10g Although it would take a lot more analysis to be sure, this move might be an inaccuracy by Gold. The move that Cyborg_sharp was more worried about on 9s (playing into the thin lines mentioned above) was a move like Eg5s Mb4n Rf1w Ra2n, defending the Hf3 temporarily by phalanx rather than by pushing the m. The obvious silver refutation is to flip the H out to unbalance Gold's army, but tactically this runs into some unclear complexities that Sharp often seems to say are good for Gold. Example line: 10g Eg5s Mb4n Rf1w Ra2n 10s ee3n Hf3w He3w ee4s 11g mf4s Eg4w Re1e Rg1e 11s hg3e Rg2n Cf2e mf3s 12g Mb5e hb6s Hd3n Ha6e ...and it's not obvious that Silver can race effectively enough around f3 or extract fast enough to prevent damage at c3 and c6. And if Silver can't play the horse flip, then Gold gains tempo because 11g Eg4w hg3n Hf3e * is a much faster way of regaining g3 than in the game if the mf4 leaves and crosses west where it really wants to go. The move also slows down mf4 from crossing west by freezing it. Maybe there's a straightforward refutation to this type of 10g here that I'm overlooking. If there is, then Silver's probably already winning after 9s. |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:12pm 10s Continuing and aiming for the camel cross anticipated on the last move. Cyborg analysis seriously considered just winning the R, but from a post-game understanding it's clear that the real threat is mc6. A log of cyborg time was spent analyzing the various threats and flips of hb6 to threaten it in c3, as well as analyzing the horse frame. Among other things, we (correctly, I think) judged on this move that mH hostage for H frame would be slightly good for Silver, even with the buried hb7, although at the time we weren't sure about that judgment. Also on this move we became far more systematic about the cyborg search. Rather than searching interactively so much, I took full advantage of the fact that computers can run 24/7 and established a schedule of twice a day uploading to Sharp a set of positions to be searched for 12 hours each, one per CPU core (reaching s18 rather than the interactive depth of s13). This practice remained effective for the rest of the game, allowing deeper search with much less human time spent. Cyborg eval: +0.4 11g The expected main line. 11s Following through! Rejected the R capture after analysis because then Gold starts pulling hb6->c3, and we end up with HR for h. After that, despite being up material the position is unclear because it's hard to deal with the h frame, and it even takes work to prevent an EM rotation. And also Sharp becomes useless in that kind of position, which is really bad for cyborg strength. Crossing immediately gives up the R, but avoids the Hh trade, and then because the M has to defend the hostage, we get the desired good-for-silver Mh hostage for H frame. Cyborg eval: +0.5 12g Expected. 12s Moved quickly, as this is pretty much forced now. 13g Rg1e was unexpected and perhaps a small inaccuracy? It's a wasted step until it's needed, and it's not needed right now. Gold paid for it on 15g and later in the game when c3 became depleted, which could have been avoided by Dd2n now instead plus one or two more steps on the next move to set up for more efficient advancement. 13s During the game we didn't look at Rb5w Mb6s Hb7s Rc7w at all (from a brief glance through the mob threads, although I didn't read anything in detail). It looks potentially strong but needs a lot more analysis - a brief interactive deepen on it shows Sharp fluctuating wildly between good for Gold and good for Silver. The chosen move was more plain and simple and maybe misses that opportunity. But given that the hostage-for-frame is good for Silver, it's still okay for Silver to go with a slower plan. So in game the plan was to pull Rh5 back to Rg5, capture it, and meanwhile free hb7. If Gold insists on holding the frame, Silver can easily break it given enough time. It's hard for Gold to gain control of c6 without giving up the frame, and hb7 alone can dismantle the frame down the h file, or two little pieces can threaten to dismantle it down the h file. (Most of this was observed and planned out on 11s and 12s). And as we discovered in analysis on later moves, advancing dogs right down the center is even better. One final detail: rc8e was a very deliberate 4th step choice to allow a future hb7n hb8e rc7w hc8s. This paid perfectly off on 15s. Overall I'm very happy with Cyborg_sharp's 4th-step choices in this game. Cyborg eval: +0.6 14g Cyborg expected Dd2n Cc2n Rc1n Dd3n instead. Sharp doesn't like the game move for Gold, although I'm less sure. Overall, the cyborg expected a D+C advance instead of spending valuable time with an R advance, with the D interfering temporarily with the Rh5/Rg5 capture to gain momentum in swarming c6. With hindsight, advancing a second R over advancing minor pieces more quickly does seem non-ideal - the c4 rabbit was quite the liability on 20g (and a little on 18g on c5). But I'm not sure about this. 14s The obvious move. Spent a while trying to decide between the game move df5e dg5w Rh5w rh6s and the immediate horse rotation hb7n hb8e rc7w hc8s. Eval jumped up again this move compared to last move. For Gold, Sharp didn't like how Gold failed to advance Dd2, which made it very awkward for Gold to (temporarily) defend Rg5 next on 15g. Cyborg eval: +1.0 15g Nice move. 15s The natural response to a dog advance, foreseen back on 13s. Cyborg eval: +1.0 16g With deep search, Cyborg_sharp seems to prefer Rc1n to De5w for Gold. Not entirely sure on the tactical details, but on 16s Silver was pretty glad to be able to play Df5w, opening up additional space. Many of the lines searched from 15s and 16s showed a silver dog down the center as a strong plan, so Gold should want to delay that possibility for a bit longer. 16s Most of the time spent this move was to decide whether Silver needed to stuff c6 or whether rd8w was sufficient defense. Cyborg eval: +1.3 |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:13pm 17g We didn't expect Gold to cross so soon! We expected the elephant cross soon as a significant possibility in analysis, but only after the Gold camel established itself on c7. Example line (each move labeled with Sharp's eval for Silver with a half-day search after playing that move): 17g cd7e Md6n hc7s Md7w +1.6 camel pushes in 17s Rg5w cg6s hc6s de6w +1.3 one of many possible natural responses 18g Cd2w Ef4w Ee4w Ed4w +1.4 cross! 18s hf3n Cf2n ee3w re8w +1.3 19g Ec4e hc5s Rb1n Rb2n obviously there are other ways Gold and Silver can play too Again, with the risk of overfitting to hindsight or simply being wrong about the tactics, I'd speculate that Gold's immediate cross is bad without first establishing Mc7. Gold had to play Mc7 next anyways in the game, and it's better to do it instead while the frame is intact. Silver can use 17s far more effectively to deal with the cross when the frame is released early, because Silver gets to spend those steps on the E and the h crossing over rather than on minor pieces in the home area. 17s There was an interesting choice this turn between: 17s hc7s rc8s hf3n ee3n (game move) 17s hc7s rc8s hf3n Cf2n (alternative) Cf2n is very non-human and it seems intuitively strange to delay the critical movement of the silver elephant. However, Cf2n does have tactical value for a future C threat in f6, and it may just turn out that Silver can do everything he needs to do with the game move except with Cf2n effectively replacing what would have been a less-useful 4th step on a future move. (Indeed, analysis was pointing towards this being the case). I still don't know which of the two is better. Gold's immediate cross resulted in another tick up in the cyborg eval. Over the next many moves the game proceeded very similarly to our analyzed lines, including the various space-denial phalanxing tactics occurring later. A snippet of the analysis tree: 17s hc7s rc8s hf3n ee3n (i) +1.7 18g cd7n Md6n rc7n Md7w (ih) +2.1 Mpush 18s cd8s ee4w de6w hf4n (iha) +1.5 19g Rc1n De2n Ec4w Eb4n (ihaa) +2.0 19s Rc5s hc6s dd6w re8w (ihaaa) +1.8 20g Eb5s mb6s Ha6e ra7s (ihaaaa) +1.6 20s dc6e Hb6e mb5n de5n (ihaaaaa) +1.7 21g Cd2n Hg3n Rg2n Eb4n (ihaaaaaa) +2.1 21s Rc4s ed4w rh5s rh6s 21g Eb4s Eb3e Rc4w Ec3n (ihaaaaab) +1.6 21s Dd5e ed4n hc5w rf7e 20g cd7s Mc7e Cd2n Cd3w (ihaaab) +2.0 20s cd6e dc6n ce6n de5n (ihaaaba) +1.6 21g hc5n Eb5e rd8e Md7n 20s de5s Dd5e ed4n de4w 18s cd8s ee4w de6w re8w (ihb) +1.5 19g Rc1n Cd2n Ec4w Eb4n (ihba) +2.0 19s Rc5s hc6s dd6w hf4n (ihbaa) +1.8 20g Eb5s mb6s Ha6e ra7s (ihbaaa) +2.0 20s dc6e Hb6e mb5n de5n (ihbaaaa) +1.8 21g De2n Hg3n Rg2n Eb4n 20s rc8w rb7w Hb6n mb5n Cyborg eval: +1.8 18g Expected and forced, with a choice of cat push. I don't think the choice of push matters much. 18s Exactly as planned on 17s. I enjoyed Sharp's tactics on this and the next several moves - the central dog basket, and the phalanxes denying space to the Gold elephant. With a concerted effort, I don't think this would have been out of reach for Gold to anticipate back on 16g and 17g, when there was still choice about how to maneuver to swarm to prepare for the E cross. Gold should have had ways to avoid this. During the game, I was quite nervous about these tactics - they seemed simultaneously very complicated and almost too good to be true. I kept wondering if Gold would find some unexpected resource and turn things around, since Sharp hadn't actually cashed in on any material yet! Despite the tactics, on this move and 20s, I think human assistance was of significant value. There were a few pitfalls, such as lines where the silver e got decentralized on to the b file to preserve the mH hostage, that Sharp didn't overtly dislike and that we carefully steered around. Cyborg eval: +1.9 19g A slightly different choice of advancing steps, but otherwise good and expected. 19s Pretty much as planned. Again, I found Sharp's ideas here to be really cool. Gold's own c4 rabbit is a perfect piece for Silver to plug everything up and deny the E all of its key squares. Spent a lot of time trying to consider alternatives and solve the tactics on this move, to ensure that Gold had no unexpected resources. This was very hard - Gold has a boggling number of possibilities in the upcoming moves. During the game we didn't feel strongly about Re8w over other 4th steps, and 4th-step selection isn't Sharp's strong point either. But I'm very happy with it now - it turns out to be the perfect step to maximize the power of both 20s and 21s. Cyborg eval: +1.8 |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:13pm 20g One of the analyzed lines, and a reasonably okay one. Other possibilities (some of which we were slightly more worried about, actually) included: 20g Hg3w Hf3n Cd2n De2n (advance and get the other H involved) 20g Eb5s mb6s Eb4w mb5s (complex tactics here) 20g cd7s Mc7e Cd2n De2n (camel escape, f6 is a bit fragile with the R not yet captured) 20s Here Re8w proved its value: 20s rc8w rb7w Hb6n mb5n (game move, only possible with Re8w) 20s ra8e rb7w Hb6n mb5n (alternative that we had analyzed more earlier) Here we analyzed and rejected the H frame. As mentioned earlier on 18s, this was another move where human assistance was useful - Sharp seemed to be overvaluing the frame, as analysis showed no clear way to forcibly convert it into material. Cyborg eval: +2.0 21g Unexpected move! This move was a gap in our analysis. Given the depth of search it took on 21s to confidently refute it, it's plausibly a good move and in another world could have gained back part of the advantage from Silver. Unfortunately, it turns out to be a blunder. Gold's best bet might be something like this, to advance and untangle the huge problem of Rc4, stubbornly hanging on and waiting for a Silver mistake. 21g Cd2n Eb4n De2n Cd3w 21s The position is good for Silver now because Gold has given up the key square of c7 for almost nothing in return. Silver has several ways to proceed now that are all much better than the anticipated +2.0 on 20s, all reaching +2.5 or higher. The two simplest are: 21s Rc4w ed4w dc6n mb6e 21s dc6e mb6e mc6n dd6w Both of them immediately capitalizing on Gold relinquishing c7 by ensuring that Gold can't get it back. Another is more aggressive: 21s Rc4w ed4w de5s de4w Simply taking advantage of the gain in time to attack and open up another can of tactics. But I think the sharpest and by far the most exciting way to play is the game move: 21s dc6n mb6e ra6e ra7s Blocking out Silver defenses and forcing a dog capture soon. The reason that this last move is sharp and exciting is because there a Gold goal attack hidden in the works if Silver follows through with the fastest capture! This is the key move: 22g Mb7w Ma7e ra6n Ra5n rabbit burrow Then if Silver takes the dog with 22s Dd5n hc5e mc6s Dd6w Dc6x, we have: 23g dc7s Mb7e * * (two advancing steps around c3, these do affect Silver's counterthreat ability, Gold has many choices) 23s Rc4s ed4w * * (no further useful steps in the west, Sharp suggests things like Rg5w hg4n) 24g Mc7w Mb7e ra7e Ra6n dig in deeper! 24s rb6w dc6w mc5n ec4n 25g cd7s Mc7e rb7e Hb8s attack A1 25g cd7s Mc7e rc8s Hb8e attack A2 24g cd7s Mc7e Md7w rd8s more use of Silver's rabbits against him 24s cd6e dc6e mc5n ec4n 25g Mc7w rc8s Hb8e Hc8e attack B1 25s Rf5n Rf6x hg5w rd7e ra8e Sharp's PV at interactive depths, and it usually loses in 5! 25s Rf5n Rf6x hg5w ce6e ra8e Although this seems to hold. 26g ra7n Mb7w Ma7e Ra6n 26s rd7e rc7e mc6n dd6w 25g Mc7w Mb7e ra7e Ra6n attack B2 25g Hb8s ra8e ra7n Hb7w attack B3 Obviously Silver can also decline to capture on 22s, but this is boring and loses a lot of time. With Sharp backing me up, I strongly suspect I would have taken on the goal attack head-on, as a way to finish the game with a bang. Cyborg eval: +3.1 22g Sadly, Gold didn't see the goal attack possibility, or maybe they did see it but thought that Silver would decline, and didn't judge it worth trying. Gold's actual move neither does a good job of prolonging and resisting nor of escalating and seeking last-ditch complications. 22s And now the game is over. When Gold resigned, we were in the middle of choosing between: 22s Rc4s ed4w dc7w mc6n 22s Rc4s ed4w dc7w cd7w Both of which seem free of any reasonable counterplay by Gold, and both easily enough to close out the game. Cyborg eval: +3.6 |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by lightvector on May 20th, 2016, 11:20pm Again, thanks to everyone. It was an interesting game! |
||
|
Title: Re: Cyborg_sharp v Mob game notes Post by Lion on May 25th, 2016, 6:29pm Thanks lightvector! I regret not being around for the initial 10-12 or so moves. Those were the crucial moves which decided what type of position we would be playing--the type where humans are strong or the type good for computer analysis. Unfortunately for the Mob, the later position arose. Still, it was a fun game, and I hope there will be a rematch someday. |
||
|
Arimaa Forum » Powered by YaBB 1 Gold - SP 1.3.1! YaBB © 2000-2003. All Rights Reserved. |