Over the last couple of months, I have been experimenting with the use of SAT solvers. This research recently came to fruition, with Tomas Rokicki and me announcing a new discovery with the intention of writing up the methodology in a paper.

Academic papers, however, typically feature a polished proof which omits all of the insights and attempts and failures which led to its discovery. This post attempts to remedy that by presenting a chronological itemisation of the journey; the technical details will be largely reserved for the paper.

### 30th January: graph theory supervisions

The previous term I had been supervising a lecture course on Graph Theory for the university. One of the questions on the final example sheet was the following:

11. Let G be a graph in which every edge is in a unique triangle and every non-edge is a diagonal of a unique 4-cycle. Show that |G| ∈ {3, 9, 99, 243, 6273, 494019}.

The usual method of solving this problem is to show that the graph must be regular, and therefore strongly regular; the resulting finite set of possibilities drops out from considering the eigenvalues of the adjacency matrix and observing that their multiplicities must be integers.

It is interesting to ask whether there actually exist suitable graphs with these numbers of vertices. For 3 vertices, a triangle obviously works; for 9 vertices, the toroidal 3-by-3 grid graph is a solution. There is also a known example with 243 vertices, related to the remarkable ternary Golay code. For 99, 6273, or 494019 vertices, however, it is unknown whether there is such a graph.

I mentioned in the supervision that John Conway had offered $1000 for a resolution of whether the 99-vertex graph exists, still unclaimed. One of my supervisees asked whether it could be amenable to computer search, and I responded that it might be within reach of a *SAT solver*: an algorithm for determining whether a set of Boolean equations has a solution.

The 99-graph problem is of a similar size to the Boolean Pythagorean Triples Problem, recently defeated by Marijn Heule and an army of SAT solvers. The UNSAT certificate (proof generated by the machine) is 200 terabytes in length, making it the largest proof of any mathematical problem in history. (The 13-gigabyte Erdös discrepancy proof, also found by SAT solvers, pales in comparison!)

At this point, I thought I would have a go at attacking the 99-graph problem myself. I’d done roughly as much exploration of the necessary structure of the graph as could be accomplished within a single side of A4 paper, and felt it was in a form amenable to SAT solving. I read several academic papers about SAT solvers, and closely monitored the competition results, to reach the conclusion that the state-of-the-art technique for approaching large SAT problems proceeds along lines similar to the following:

- Initially simplify the problem using
**lingeling -S**; - Split the problem space using
**march-cc**into thousands of separate ‘cubes’; - Share the cubes amongst parallel instances of the incremental SAT solver
**iglucose**.

I started to develop a framework, which I called MetaSAT, for automating this process. Coincidentally, at the same time, someone else also became interested in SAT solvers…

### 26th January: Oscar Cunningham’s LLS

In his Odyssean epic *The Art of Computer Programming*, Donald Knuth dedicates a large section to SAT solvers with a particular emphasis on applying them to Conway’s Game of Life. This left the realms of theory and entered practice when Oscar Cunningham, who studied mathematics in Cambridge at roughly the same time as I did (and taught me how to play frisbee), wrote a Python script to search for patterns using Knuth’s ideas.

*Logic Life Search*, or LLS, is a sophisticated Python search program capable of using one of many different SAT solvers as backends to find patterns in various cellular automata.

This drove me to attempt to integrate these ideas into the MetaSAT project, so that I could search larger spaces than is possible with a single SAT solver. What I ended up doing was creating a simplified Golly-compatible version of LLS, called GRILLS. Whilst LLS is strictly more powerful, GRILLS has a more user-friendly graphical interface, and allows for loading results from previous runs:

I departed slightly from Donald Knuth’s binary-tree approach of counting neighbours, instead opting for a ‘split-radix’ binary-ternary tree. This reduced the number of variables needed by the SAT solver by 40%, without affecting the number of clauses, thereby making the problem slightly easier to digest (at least in principle).

The pinnacle of its success was being able to find an already-known c/4 orthogonal spaceship. Whilst showing that SAT solvers could replicate existing discoveries, there was no evidence that they could find anything new — in other words, they seemed to be trailing behind discoveries found with special-purpose CA search programs.

### 31st January: Tomas Rokicki’s c/3 spaceship

This changed at the end of the month. Tomas Rokicki had also implemented a Perl script to search for spaceships in cellular automata, and came across a new discovery. It is tiny, completely asymmetric, fits in a 13-by-13 box, and remains to be named anything other than its systematic name of xq3_co410jcsg448gz07c4klmljkcdzw71066.

He then turned to attempting to find *knightships*: elusive ships which move parallel to a knight in chess, rather than orthogonally or diagonally. Prior to 2010, oblique spaceships had never been built, and the only examples since then were huge slow constructions. The smallest two, Chris Cain’s *parallel HBK* and Brett Berger’s *waterbear*, are approximately 30000 cells in length. The waterbear is the only one which moves at a respectable speed, and its creator produced a video of it:

Finding a small knightship has been considered one of the holy grails of cellular automata. The closest attempt was Eugene Langvagen’s 2004 near-miss, which has a slight defect causing its tail to burn up:

In six generations, it almost translates itself by (2, 1), but differs in two cells from its original state. Very soon, this error propagates throughout the entire pattern and destroys it. In the 14 years following that, no-one has been able to find a perfect knightship.

Tom’s search, however, did yield increasingly long and increasingly promising partials. He found four further 2-cells-away near-misses, but no 1-cell-away or perfect solution, and statistical analysis (modelling the distribution of errors as a Poisson distribution) suggested that we would need another 2 years of searching to find a result.

### 9th February: a phoenix from the ashes of pessimism

On 3rd February, Oscar Cunningham cast doubts on the efficacy of his search program for finding a knightship, remarking the following:

I’m under the impression that programs called -find are just much better at finding ships than programs called -search. Is that true? If so then there’s not much point looking for a knightship with LLS.

He was referring to David Eppstein’s *gfind* search program, which conducts a depth-first row-by-row brute-force backtracking search. This, and its descendants, have been responsible for the vast majority of new spaceships discovered in cellular automata. David Eppstein wrote a paper detailing the methodology, which I suggest reading before proceeding with this post. (If you’re impatient, then take a cursory glance now and read the paper properly later.)

This gave me an idea: what if somehow Tom Rokicki’s partial-extending approach could be hybridised with the row-by-row tree-searching approach in gfind? I sent an e-mail to Tom to this effect:

Good news: I’ve realised that the methodologies of knightsearch and gfind

are entirely compatible. In particular:(a) Represent the search pattern as a two-dimensional lattice of unknown

cells by interleaving rows of different generations. (The way to do this for

a (2, 1)c/6 knightship is to take Z^3 and quotient by (2, 1, 6), giving a

two-dimensional lattice of unknown cells. Unlike in gfind, the ‘rows’ are

exactly perpendicular to the direction of travel, even for knightships.)Specifically, we map cell (x, y, t) to lattice coordinates (2x-y, 3y+t).

(b) The ‘state’ of the search is a 6-tuple of consecutive rows of length W.

The ‘search grid’ is a width-W height-(K+7) grid of lattice points, where

the first 6 rows are populated by the state, the next row is the ‘target’,

and the following K rows are for lookahead purposes. There are a few columns

of forced-dead cells on either side of the search grid. Pass this problem to

an incremental SAT solver to find all possibilities for the 7th row which

are consistent with the remaining K rows (I think gfind uses K = 1 or 2; SAT

solvers can clearly go much deeper). Each time the incremental SAT solver

gives a solution, provide the clause given by the negation of the 7th row

so that the next run gives a new possible 7th row. Repeat until UNSAT. The

‘child states’ are the 6-tuples obtained by taking the last 5 elements of

the parent 6-tuple and appending one of the possible 7th rows.(c) Run a breadth-first search starting with a small partial, using (b) each

time a new 6-tuple appears. We can canonise 6-tuples by doing a horizontal

shift to centre the non-zero columns into the middle of the search grid.

This means that the knightship needs only be *locally* thin, and is thereby

permitted to drift in various directions (similar to in knightsearch). The

narrowness of [the bitwise disjunction of the members of] a 6-tuple may

be a good heuristic to decide which nodes to attempt to extend first.(d) If you hit a dead-end, you can always rerun the already-found tuples

using higher values of W. Hence, the width need not be known in advance

when commencing the search. (If W increases particularly high, then it

may be necessary to decrease K to avoid overloading the SAT solver.)

Within the next few days, I implemented the aforementioned search program. It used the same backend (iglucose with named pipes for communication) as Tom’s search, but was otherwise very different.

### 19th February: adaptive widening

Over the coming weeks, the search program grew in complexity, gaining various features and heuristics. In programs such as gfind, the search-width is set as a constant, and it either finds a solution or exhausts the search space to no avail. I implemented *adaptive widening*, where the latter causes the width to increment and the completed search to act as a backbone for the new (higher width) search.

Another approach, borrowed from cryptanalysis, is meet-in-the-middle: separately grow search trees of fronts and backs, and compare them to see whether they can join to form a complete solution.

### 25th February: save-and-restore

One of the later changes I made, which logically should have been the first feature, is to save progress to disk and restore it. That way, I no longer needed to start the search afresh whenever I added a new optimisation to the program. I spent the next week running a width-35 search, saving the tree to disk every hour.

### 6th March: a last attempt in desperation

I resumed the search at width-0 to allow adaptive widening to take over, as smaller search widths lead to faster SAT solving. I also decided to adjoin one of Tom Rokicki’s partial results to the search tree to make things more interesting: this way, it would be able to piggy-back from his independent research. I watched the search width gradually increase as each completed without success, and as the night was drawing to a close it looked as though width-24 would finish as the queue gradually emptied:

Increasing head search width to 24... 280700 head edges traversed (qsize ~= 47966). 280800 head edges traversed (qsize ~= 47616). 280900 head edges traversed (qsize ~= 47369). 281000 head edges traversed (qsize ~= 47200). 281100 head edges traversed (qsize ~= 46993). 281200 head edges traversed (qsize ~= 46516). 281300 head edges traversed (qsize ~= 46180). 281400 head edges traversed (qsize ~= 45471). 281500 head edges traversed (qsize ~= 44632). 281600 head edges traversed (qsize ~= 44364). 281700 head edges traversed (qsize ~= 44036). 281800 head edges traversed (qsize ~= 43808). 281900 head edges traversed (qsize ~= 43231). 282000 head edges traversed (qsize ~= 43021). 282100 head edges traversed (qsize ~= 42759). 282200 head edges traversed (qsize ~= 42272). 282300 head edges traversed (qsize ~= 41795). 282400 head edges traversed (qsize ~= 41256). 282500 head edges traversed (qsize ~= 40706). 282600 head edges traversed (qsize ~= 40428). 282700 head edges traversed (qsize ~= 39523). 282800 head edges traversed (qsize ~= 38931). Saving backup backup_head_odd.pkl.gz Saved backup backup_head_odd.pkl.gz Saving backup backup_tail_even.pkl.gz Saved backup backup_tail_even.pkl.gz Backup process took 53.0 seconds. 282900 head edges traversed (qsize ~= 38000). 283000 head edges traversed (qsize ~= 37268). 283100 head edges traversed (qsize ~= 37023). 283200 head edges traversed (qsize ~= 36598). 283300 head edges traversed (qsize ~= 35869). Increasing tail search width to 11... ...adaptive widening completed. 283400 head edges traversed (qsize ~= 35024). 283500 head edges traversed (qsize ~= 33950). 283600 head edges traversed (qsize ~= 32596). 283700 head edges traversed (qsize ~= 31262). 283800 head edges traversed (qsize ~= 29728). 527600 tail edges traversed (qsize ~= 149949). 283900 head edges traversed (qsize ~= 28707). 284000 head edges traversed (qsize ~= 26488). 284100 head edges traversed (qsize ~= 24744). Saving backup backup_head_even.pkl.gz Saved backup backup_head_even.pkl.gz Saving backup backup_tail_odd.pkl.gz Saved backup backup_tail_odd.pkl.gz Backup process took 61.7 seconds. 284200 head edges traversed (qsize ~= 21792). 284300 head edges traversed (qsize ~= 19753). 284400 head edges traversed (qsize ~= 17345). 284500 head edges traversed (qsize ~= 14687). 284600 head edges traversed (qsize ~= 14041). 527700 tail edges traversed (qsize ~= 128776). 284700 head edges traversed (qsize ~= 12097). 284800 head edges traversed (qsize ~= 9503). 284900 head edges traversed (qsize ~= 8984). 285000 head edges traversed (qsize ~= 8388). 285100 head edges traversed (qsize ~= 6719). 285200 head edges traversed (qsize ~= 6670). 285300 head edges traversed (qsize ~= 6606). 285400 head edges traversed (qsize ~= 6548). 285500 head edges traversed (qsize ~= 6409). 285600 head edges traversed (qsize ~= 6463). 285700 head edges traversed (qsize ~= 6405). 285800 head edges traversed (qsize ~= 5271). 285900 head edges traversed (qsize ~= 5047). 286000 head edges traversed (qsize ~= 5030). 286100 head edges traversed (qsize ~= 5027). 286200 head edges traversed (qsize ~= 5089). 286300 head edges traversed (qsize ~= 5066). 286400 head edges traversed (qsize ~= 5056). 286500 head edges traversed (qsize ~= 5142). 286600 head edges traversed (qsize ~= 5241). 286700 head edges traversed (qsize ~= 5326). 286800 head edges traversed (qsize ~= 5344). 286900 head edges traversed (qsize ~= 5394). 287000 head edges traversed (qsize ~= 5298). 287100 head edges traversed (qsize ~= 4950). 287200 head edges traversed (qsize ~= 4618). 287300 head edges traversed (qsize ~= 4601). 287400 head edges traversed (qsize ~= 4646).

But then, something remarkable happened. It spewed out a series of long partials, which had extended from about two-thirds of the way along Tom Rokicki’s input partial and overtaken it! I then saw something which caught my eye:

Found partial head of length 249. As plaintext: .......................................* .....................................*** ....................................***** ...................................******.* ...................................***.*..* .................................**.*..*.* .................................**....*** ................................*.**....*.** ................................*....*.*..** .................................*...*.**..*.* .................................*...****** ................................*.......*.*.*...*.* ..............................***..........*....*.** .............................****.*...*.*.*..*.***** .............................*******.*.*..****..*.*** ...........................**.*.*.*....*....*.*.**..* ...........................**...*.*......*..***.**** .........................****..**.**........**...*.* ..........................**.......*..*...*.**......* ........................***........**.*.*.*.*....*.*.* .........................*..*.....**........*....**.** .......................**...*.*.*..*.......**..*..*** .......................*.**.*.*.**.*.*....***......*** .....................*..*.***.*.*.........*.***.....* ......................*......*..*.*.......**.***.*** .....................**.*.......*.*.....**** ....................*..*.*.*..**....*.*..*.**.....* ....................*.***.*...**....*.**....*...* .....................***.*....**.......*.*.....* .....................*...*.*.*..*.**.*..*...**.* ......................**.**...**..*.....*..***.*.* ........................*...*..***.**...**.**.*..* .......................**.*...**.**.*...**.*....* ......................*.*.......*..*...**..*.*..* ......................*.*..........*.*..*..*.*....* ................................*..*.*..**.*.**...* ..............................*.*......*.*.***.*..*.* ................................*..........*...*.*..* ..................................*..*.*.*....**.*.* ..................................*.*.*....*.***.*.** ....................................*.*..*.*...*....** ....................................*.**.***.....*.*.* ....................................*....***.*...*..** .....................................****..*.*.*..** .......................................*...***.**.**.* ...................................**..*...*.* ..................................***..*...**....*** .................................***...**......*.*.* ................................****.**.**......*.** ................................***..**....*.*..* ...............................****..*..*..*..*** ................................*....*....***** ................................*.....*...**..*..* ..............................*.*..*.******......* .................................*.**.*...**.**..* ...............................*.*.**.*...*...*.** ...............................*....***...**...** ................................*....*....*****.** ................................**..**..*..*.*..* ..............................****..**..***..*..*.* ..............................*.**..**.....*....* ..............................*.*....*.**..*...** .................................*..**.*.* ..............................******.**..*.*.** .............................***...*.*.*.*.****.* .............................*..............***** ...............................*.*....*.....**.** ................................*.*.*.*.*...* ..............................*.*...........**.* ..............................*.*.*.......***..** ............................*..............*...*** ...............................*...*.*.....*...*.* .............................*.*.*.......*.**.*..* ...........................*..*....*......*...*.** ...........................*.*............**.....** ...........................***.*..*.......**..*.** ............................**..........*.******* ..........................****.**.........***.*.* .............................*..................* ...........................***.*..............*.*.* ...........................*.*.......*.*.*.*.* ............................*......***.........* ..................................****.......* ..........................*......*****.......* ........................*........***......* ........................*..*.*.**.*..* ........................**.*..*......* ......................******..*.***..* ......................**.****.**.**..* .....................*.**..*......*.* .....................*.*...**.....*.*.* ...................*..*....**.*.**....* ...................*.***.****....*..*.** ....................***....*..*..*....** ....................*.........*.**..*.*** ........................**.*...*.*.**** .......................*.*.*.*...*..****...** ......................**.***.*...**.......**** .....................*.*.............*..****** .....................***.*....*.**..*******.** .....................**..**.*.*.*........**.** ...................*.**..*.*....*.**.....**.* ....................*.*.............**..*...* ....................*...*........**.*...* ...........................*.*.*.*..*...*.* .............................*....*.*...*.* ..................................*..*..*.* ................................*.**** ............................*......***.* .................................*.* .................................*.........* .......................................**..*.* .................................**.*.***** ................................***********.* ...............................**.*.*.*..*.* ................................*.*...*..*.** ..............................*.*.*.***..**** .............................**......*....*.*** ............................***.*........*...** ............................***.*.......*.**.* .............................**.*.*...*.****** .............................**.***......*...** ..............................*.**...........* ..............................*.***..*.*.*.* ..........................*.......*.......* ..........................*.*...*****.....* ........................**********..*.*.*.*.* ........................***.*...**.**.......* ......................*****......*.*...*.* .......................**....*......*...*** .....................***............**..*.*.* ......................*...*....***..**.***.** .....................**..****..*.*.*.*..**** ....................*.**.**.*.*..*....*.**.** ....................*..*.**.****.*.***..*....* .....................*.......*......*..*...*.* .......................**.*.***.....**.*..** ......................*.*....*.**....***..* .........................*..*..*....**.* ......................*.*..***.......* ......................**....** .....................*.*.**.* .......................*** ........................*..* .........................*.*.* ....................*..*.**..* ....................**.*.** ....................**.*.* ....................**.*.* ..................**.*.*.*.* ..................**.*.*.*** .................***.*.*.** .................*...*.*.* ................*..*...*.* ................*....*....* ...............*.*...*....* ...............*.*...****.* .............*****..****.*..* .............**.*...*.*....* ...........***.*......*.**.* ...........**..**.*.*.*.*..* ..........***..***...*..** ..........***..*..*...*.* ..........**..........* ..........*....*..**.** ........*.*..*.*.......* ...........*.***.*.*.*.* .........*.*.*.*******.* .......*.*...*.***...*.* .....*....*....*.*.*.*..* ........*............*.** ......*.**.*.*.*.***.*.*..* .........*.******.**.**** ......***.*.**.......***.* ......**.*........*..*..*..* .....*.****........*....* .......***.*.*.*...*...**.* .............*....**...* ...........*.*.*.***..**.* ......*..........*.***.*.* ......*...*.*.....*****....* ..........*.*...*.......** ..........*........*.....** ................*******.* ................**...**** .............*....*.....* ...........*............*.* ..............*.****.**.*.* ............**.*.***.*.*..* ...........*.*.........* ..........****.........* ........*...*.*...* .........*.**.....*...* .......***.**.*.*.*...** ......***...*.....*...** .....****......*.*..***** .....*.*....*.**.*..**.** .....***...*....*.*.**.** .....****..*.*.**...**.* .....*....**.*.**..**..* .......*....*..**..* .....*.***..*..*...* ......*.*..**..*..**.* ......*.**.*...*..**.** ......*...*.*..*..**.* .........**..**....*.* .......****..**..*** .....**..*...**...*.* .........*..**** .......***..*****.*.* .....*......**...*..* ...*.*.**.....*..* ...*..****.....** ....**.*.......** ..*.***....*.*..* .....*.***.*...* ...**.*.*.**...*.* ...***....*..***.* .*..*.*.*.**..*..* ....*....*.*.* ....***...*.** ..*.......****.* .....*.****.* .....*.*......* .........*..*.* .........*.* ....*......* ....****.*.* ..***.*......* ..*.*.*.* .****.*.* .**.......* *** .** .**.** ..**.......*.* .*.........**** ..........***** ........**.*.*** .......***..*..* ......****.**.** ......***...*.* ....**.*......** ....** ...*.**.***.*.* ...*..*.*.*.* ....*.....***..* ....*.*...*.* .......*..*.*.* .....*....**..* ...*.*....**.** ........**...**......* ....*..***...***..*.** ....**...*..**.*.**.*** ..****.......***.***.** ..*.**.***....*...**.** ..*.**.*.*.*.....***.** ...*.*.**.**.**..**...** .*.....*.***.*.....*..* .*.***.*..*......*..*..* ....*.....*....*.*.* .....*.*.**.*.....*.*.* ..*...........*....*.*.* *.**.***........*****..* *******.***...****..*.* *...***.....*...**.*..*

Near the bottom of the terminal screen (highlighted in red above), I saw that the partial solution had a slanted fissure! It looked as though one partial had completed and another had begun immediately below and to the right of it! It seemed too good to be true, but I tried running it through ikpx2golly to convert it into a runnable Life pattern. Sure enough, the back partial burned away to leave, emerging unscathed from the ash, a perfectly self-contained knightship:

I checked the stdout logs of the search program, and it transpires that it had also found the completion; it just so happened that it was hidden under the barrage of following pseudo-partials that I didn’t notice it until several minutes later.

### 7th March: Hello, world!

The first thing to calculate was the attribution: since it grew from Tom Rokicki’s partial, there was the issue of determining who ‘discovered’ it. Fortunately, the row-by-row approach used by the search program made this exactly calculable: Tom found the front 62% of the pattern, and I found the rear 38%.

Nomenclature was the next thing to consider. Again, the knightship has a long systematic name, but something more catchy was in order. The name ‘seahorse’ was suggested by Dave Greene as a clever pun; the pattern resembles a seahorse, and ‘horse’ is an informal term for a knight in a game of chess. A pseudonymous user named Apple Bottom suggested the alternative name of Sir Robin, after the character in Monty Python and the Holy Grail. As this discovery is a knightship, considered a ‘holy grail’ in CA searching, and was found by a program written in Python, that name stuck.

Even though I wanted to hold back the publication until this article was published, it was beyond my control at this point. The discovery soon found its way onto Hacker News where it accumulated 727 up-votes and a horde of comments.

### 8th March: The discovery gets digested

After 48 hours, the surprise gradually settled and the cellular automata community became accustomed to this new discovery and using it as a component of larger constructions. In particular, Martin Grant found a configuration capable of eating Sir Robins, and there has been talk of trying to create a knightpuffer (arrangement of Sir Robins which leave behind sporadic debris).

Soon afterwards, on the 10th March, Dave Greene published an article on Sir Robin.

### What’s next?

It is not too difficult to prove that a pattern in Conway’s Life can translate itself by (a, b) in p generations only if |a| + |b| <= p/2. For all periods p up to 7, we now have such examples, but there is currently no known (2,1)c/7 knightship. It may be possible to find such a pattern using these search methods, but both Tom Rokicki and I are busy with a backlog of other things.

Consequently, you are just as likely as anyone else to find a period-7 knightship — 10 weeks is nearly twice as long as the total chronology of this post, and it’s unclear how many more CPU-hours and ideas would be sufficient to yield this gem.

As for SAT solvers more generally, I am investigating whether they can be applied to my own area of research (topological data analysis, which might just be combinatorial enough to be susceptible to these ideas). If not, there’s always that 99-vertex graph problem…

Looking back, it appears that Nick Gotts is the one who actually suggested the name “Sir Robin”? Regardless, I have to say that this is certainly an exciting discovery!

Pingback: Assorted news | Complex Projective 4-Space

In a website that showed eaters that erase a “copperhead” (C/10 spaceship), somebody remarked that such was a real waste: destroying a rare pattern that may not occur naturally. Same can be said HERE, of any eater that destroys the Sir Robin, a HUNDREDfold.

BTW, this 78×31 Knightship I installed gen. by gen., in my archives (Vol.27 – otherwise the least interesting of all 34 volumes).

Is there a copy available of the paper you’ve submitted to NIPS, perhaps in ArXiv?

Pingback: Shallow trees with heavy leaves | Complex Projective 4-Space