Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/tuzz/supersat

An attempt to find superpermutations by reducing the problem to SAT.
https://github.com/tuzz/supersat

Last synced: 3 months ago
JSON representation

An attempt to find superpermutations by reducing the problem to SAT.

Awesome Lists containing this project

README 

        
## Supersat



An attempt to find superpermutations via a reduction of finite state machines to the

[boolean satisfiability problem](https://en.wikipedia.org/wiki/Boolean_satisfiability_problem).



## Setup



To run the test suite:



```

cargo test

```



To run the application:



```

cargo run --release

```



## Overview



This is my third attempt at the superpermutation problem. This time I decided to

reduce the problem to SAT and run a state of the art SAT solver on the resulting

equation. This is very much inspired by my [experimental programming language](https://sentient-lang.org/)

and prior success in reducing problems to SAT.



Unfortunately, I did not manage to solve any unsolved superpermutation problems,

but I was successful in reducing the problem to SAT which is explained in detail

below. My hope is that someone with greater experience encoding problems in SAT

can take these ideas and experiment with different techniques to see if it's

possible to find superpermutations by this approach.



## Sorry this isn't a paper



I think there's more than enough interesting material here to write a paper, but

I'd like to move onto other things as I've already invested a lot of time into

this problem. I'm an independent researcher and its expensive for me commit the

additional time and energy required to this endeavour.



Instead, I'll try to explain things as best I can here.



## An observation



This project stems from the idea that we can define a language `S` of

superpermutations. We can create a finite state machine that recognises all

strings in the language so that it accepts strings that are valid

superpermutations and rejects strings that are not.



## One possible approach



A superpermutation is defined as 'a string that contains all permutations for N

symbols' and so we can construct our finite state machine by intersecting

simpler machines that recognise each of the permutations. In other words, we can

take the cartesian product of N! machines.



This is a widely studied area and there are existing tools that can do this for

us. For example [Ragel](https://en.wikipedia.org/wiki/Ragel) can generate finite

state machines from regular expressions, so for N=3 we can construct our state

machine as follows:



```ragel

// definition.rl



%%{

  machine S;



  main :=

    [123]* '123' [123]* &

    [123]* '132' [123]* &

    [123]* '213' [123]* &

    [123]* '231' [123]* &

    [123]* '312' [123]* &

    [123]* '321' [123]* ;

}%%

```



We can then turn this into a graph like so:



```sh

ragel -V definition.rl > graph.dot

dot -Tpng graph.dot > graph.png

```



This generates a rather large finite state machine that can be seen [here](images/graph-3-full.png).



The above machine will accept any string that is a superpermutation for N=3. As

it stands, this isn't very useful, but we can leverage this approach to find

minimal superpermutations (i.e. the shortest) for some N.



The way to do this is to intersect a machine that accepts strings of a given

length. If the length is shorter than the minimal superpermutation, the

resulting machine won't accept any strings (i.e. the language will be

empty). We can incrementally increase the length of the string until the result

is a machine that accepts something.



Before presenting this definition, there are some optimisations we can apply:



1. When listing the regular expressions to intersect, we start with the one that

limits the length of the string. This is more efficient because the diameter of

the resulting graph is limited by this length at each stage of intersection. If

instead we intersected this regular expression at the end, we'd effectively

generate the entire machine as previously before throwing most of it away.



2. We can insist the string begins '123' to

[break symmetries](https://en.wikipedia.org/wiki/Symmetry_breaking). In theory

this should reduce the amount of work by N! and we lose no generality by doing

this. We may as well combine this with the regular expression that limits the

length. The other regular expressions can now be changed to begin `[123]+`

instead of [Kleene star](https://en.wikipedia.org/wiki/Kleene_star) since we

know these permutation do not occur at the start of the string.



We know the shortest superpermutation for N=3 has length 9, so here's the

definition of that machine:



```

%%{

  machine S;



  main :=

       '123' [123]{6}   &



    [123]+ '132' [123]* &

    [123]+ '213' [123]* &

    [123]+ '231' [123]* &

    [123]+ '312' [123]* &

    [123]+ '321' [123]* ;

}%%

```



This generates a final state machine with 10 states that can be seen [here](images/graph-3-09.png).



The transitions are the ASCII characters for '1', '2' and '3'. The only string

this machine accepts is '123121321' which is the minimal superpermutation.



If instead, we limited the length to 8 (which would be defined as `{5}`), the

resulting machine wouldn't accept anything as seen [here](images/graph-empty.png).



If instead, we limited the length to larger values, these are the resulting

machines:

  - [graph for length 10](images/graph-3-10.png)

  - [graph for length 11](images/graph-3-11.png)

  - [graph for length 12](images/graph-3-12.png)



These machines accept strings that are superpermutations of length 10, 11 and 12

respectively. I think these graphs are quite pretty.



## Larger values of N



Here is the machine definition for the shortest superpermutation for N=4:



```

%%{

  machine S;



  main :=

      '1234' [1234]{29}    &



    [1234]+ '1243' [1234]* &

    [1234]+ '1324' [1234]* &

    [1234]+ '1342' [1234]* &

    [1234]+ '1423' [1234]* &

    [1234]+ '1432' [1234]* &

    [1234]+ '2134' [1234]* &

    [1234]+ '2143' [1234]* &

    [1234]+ '2314' [1234]* &

    [1234]+ '2341' [1234]* &

    [1234]+ '2413' [1234]* &

    [1234]+ '2431' [1234]* &

    [1234]+ '3124' [1234]* &

    [1234]+ '3142' [1234]* &

    [1234]+ '3214' [1234]* &

    [1234]+ '3241' [1234]* &

    [1234]+ '3412' [1234]* &

    [1234]+ '3421' [1234]* &

    [1234]+ '4123' [1234]* &

    [1234]+ '4132' [1234]* &

    [1234]+ '4213' [1234]* &

    [1234]+ '4231' [1234]* &

    [1234]+ '4312' [1234]* &

    [1234]+ '4321' [1234]* ;

}%%

```



Resulting in [this machine](images/graph-4-33.png) which accepts the minimal

superpermutation which is '123412314231243121342132413214321'.



Again, if we decrease the length by 1, the [resulting machine](images/graph-empty.png)

won't accept anything.



If we increase it by 1 to 34, [this is the result](images/graph-4-34.png).



For N=5, the approach breaks down and Ragel takes too long to generate the

machine. As the length increases, it takes an exponentially increasing amount of

time:



- length 50 takes ~1 second to generate

- length 60 takes ~6 seconds to generate

- length 70 takes ~50 seconds to generate

- length 80 takes ~15 minutes to generate



The shortest superpermutation for N=5 has length 153, so it's a long way from

being able to generate this in a reasonable amount of time.



There could be more optimisations, perhaps by subdividing the regular

expressions into groups, computing the cartesian products of these, then

combining at the end, but I didn't explore this avenue further. Even if we

could generate this machine for N=5, I'd be surprised if this scaled to N=6



It would interesting to do some complexity analysis on this approach to see how

it scales. Under the hood, I think Ragel uses

[Hopcroft's algorithm](https://en.wikipedia.org/wiki/DFA_minimization#Hopcroft's_algorithm)

to minimize finite state machines and the cartesian product of two machines can

result in AxB the number of states in the worst case.



## An alternative approach



Rather than trying to construct a finite state machine that is the cartesian

product of smaller machines, we could instead think about running these smaller

machines in parallel over the string. If they all reach their accepting state

then the string is a superpermutation.



Here's a machine that accepts strings containing '1234':



![graph-1234](images/graph-1234.png)



Here's a machine that accepts strings containing '1243':



![graph-1243](images/graph-1243.png)



Notice that the only difference between these machines are the transitions from

the '12' and '123' states as highlighted above. When these machines are run in

parallel, we know that whenever the '1234' machine is in state '12', so will

the '1243' machine. That means these states are non-distinguishable and can be

collapsed into one.



Therefore, a more natural way to think about running these machines in parallel

is instead to construct a non-deterministic machine that can be in multiple

states at once.



Here's a non-deterministic machine for N=3:



![graph-3-nfa](images/graph-3-nfa.png)



This machine will share as much state as possible when it runs over the string.

In general, the number of states in the first rank will be N, then N x N-1, etc.

So for N=6, there will be a total of 6 + 30 + 120 + 360 + 720 + 720 = 1956 states.



However, notice that this machine isn't quite what we want. It no longer accepts

superpermutations because, by definition, a non-deterministic machine accepts

its input if **any** final state is reached. We want our machine to accept its

input if **all** final states are reached. We'll address this shortly.



Before doing so, here are some things noteworthy about this machine:



1) This machine is essentially a [trie](https://en.wikipedia.org/wiki/Trie) of

all permutations for N



2) This machine has the useful property that it can only be in one state of each

'rank' at a time. Here a 'rank' is defined as the set of states representing

substrings of the same length, for example, `12, 13, 21, 23, 31, 32` form a rank



3) It is possible for the machine to not be in any states for a rank, except the

first rank which will always have an active state of the symbol that has just

been read from the input string



As it stands, we don't have a method of finding superpermutations. The machine

above will only accept strings that are superpermutations (provided we change

the semantics for when a string is accepted to specify that all final states are

reached). We will address this now by formulating a boolean satisfiability

problem.



## Boolean Satisfiability



This approach takes inspiration from [this paper](https://arxiv.org/pdf/1402.2184.pdf)

which successfully proved the Erdos Discrepancy Conjecture for C=2. We will

attempt to reduce the problem of finding superpermutations to SAT and run

state-of-the-art SAT solvers on the resulting equation.



Our approach will take the non-deterministic state machine detailed above and

transform this into a problem of boolean logic so that the formula is

satisfiability if and only if **all** final states are reached after the machine

has read its input.



In the following sections, the reduction will be described.



**1. Representing states**



To represent states, we make use of the observation that the machine can be in a

maximum of one state at a time in each rank. This means we can represent a rank

of states as a binary number indicating the state that is active in the rank.

Each bit in the binary number corresponds to a variable in the SAT equation.



For the case when no states are active, we'll insist the machine is in a 'dead'

state for that rank, which we'll number 0 in our binary representation. That

means for N=6, we'd need the following number of variables per rank:



```

rank 0: 3 variables         //   ceil[ log2(6)          ] = 3 bits

rank 1: 5 variables         //   ceil[ log2(30 + dead)  ] = 5 bits

rank 2: 7 variables         //   ceil[ log2(120 + dead) ] = 7 bits

rank 3: 9 variables         //   ceil[ log2(360 + dead) ] = 9 bits

rank 4: 10 variables        //   ceil[ log2(720 + dead) ] = 10 bits

rank 5: 10 variables        //   ceil[ log2(720 + dead) ] = 10 bits

```



In total, this means we need 44 variables to represent the machine's state for

N=6. If we'd taken the alternative approach of representing states as individual

variables in the SAT equation, we'd instead need 1956 variables. It is not clear

that reducing the number of variables in this way necessarily helps the SAT

solver but this is the approach taken in the Erdos paper, so we opt to do the

same here.



An interesting line of investigation would be to experiment with different state

representations and see what impact this has on various solvers.



**2. Representing time**



As the machine scans the input string, it changes state. Therefore, we need to

represent the state of the machine at each point in time. We introduce the

notion of a 'snapshot' which is the state of the machine after scanning each

character of the input string. If each snapshot contains 44 variables, then to

test a string of length 872, which is the shortest known superpermutation for

N=6 would require 44 x 872 = 38,368 variables.



**3. Representing input**



To model the input string, we don't need to add any additional clauses. Instead,

we leave the first rank of states in the machine unconstrained and treat this as

the input string. For example, if state '1' is active at time=0 then this

represents a '1' at the start of the input string.



**4. Modelling transitions**



To model the behaviour of the machine, we add clauses for the state transitions.

These are implications between states in consecutive snapshots, according to the

graph of the machine presented above. If the machine is in some state at time

t, then it may transition to the next state at t+1 if the symbol that was just

read as input matches the transition.



Here is an example clause that is added:



```

State(name=12, time=5) -> State(name=123, time=6) ^ Input(symbol=3, time=6)

```



Which is equivalent to the following:



```

State(name=12, time=5) -> State(name=123, time=6) ^ State(name=3, time=6)

```



Of course, this isn't in conjunctive normal form and a state is represented by

multiple bits. To convert this to CNF, we use equivalence rules. If A and B are

sequences of 3 bits, then `A -> B` can be written in CNF as:



```

(-a0 v -a1 v -a2 v b0) ^ (-a0 v -a1 v -a2 v b1) ^ (-a0 v -a1 v -a2 v b2)

```



We also add clauses for the implication in the opposite direction:



```

State(name=123, time=6) ^ State(name=3, time=6) -> State(name=12, time=5)

```



Strictly speaking, we only need implication in one direction, but adding both

seems to help the solver find a solution more quickly. I'd speculate this is

because it removes ambiguity from the SAT solver about whether to set the state

but this requires more analysis.



Finally, we do not add explicit transitions for the dead states, nor do we add

the 'looping' transitions on the final states. Instead, we handle these cases

differently as explained later.



**5. Modelling the dead state**



Our choice of encoding for states means that a unique number is determined for

each rank, corresponding to one state being set in each rank. We arbitrarily

number dead states as 0 but currently our binary representation does not

preclude numbers that do not correspond to states.



Therefore, we add clauses to explicitly ban all numbers that do not correspond

to states. If the numbers 0 to 12 correspond to valid states then that means we're

using 4 bits for this rank and need to ban the numbers 13 to 15. In binary, 13,

14 and 15 are represented as 1101, 1110 and 1111 respectively. For the following

clauses, we'll use a, b, c, d to denote the variables for each bit from left to

right:



```

(-a v -b v c v -d) ^      // ban 13

(-a v -b v -c v -d) ^     // ban 14

(-a v -b v -c v d)        // ban 15

```



To take the 'ban 13' clause as an example, this will be false for the assignment

`a=true, b=true, c=false, d=true` in the SAT equation, so it must choose

something else.



When we take the transition clauses and combine these with the clauses to ban

invalid numbered states, the result is the dead state is the only remaining

state available for the machine to be in.



This would be the case, for example, when the input string contains repeated

symbols that do not correspond to permutations, e.g. '12333'.



**6. Modelling the final states**



As mentioned above, we do not add 'looping' transitions on the final states. We

could do, but this would result in a lot of clauses and there's a more efficient

way to encode the requirement that each of the final states must be reached.



Our approach introduces an additional set of variables for each permutation that

we'll refer to as the 'Goal'. The intended meaning of these variables is to

represent another number in binary. Each number is the time at which the machine

is in the final state for that permutation. So, for example, if the '321'

permutation appears at time 5 in the input string then the number corresponding

to permutation '321' would be 5.



We model this in CNF by adding clauses of the form:



```

Goal(name=321, time=5) -> State(name=321, time=5)

```



We then ban all binary numbers that do not map to valid times that are beyond

the length of the input string. We use the technique described in the previous

section to do this. The result is that every permutation must occur in the input

string at least once since the goal variables must refer to a time at which the

final state is true for that permutation.



**7. Initial configuration**



We also add clauses to break symmetries so that the input string begins with

ascending numbers. This is to break symmetries in the problem domain:



```

State(name=1, time=0)

State(name=2, time=1)

State(name=3, time=2)

```



We also add clauses that insist all the dead states are initially active in the

machine. This prevents final states being reached before the necessary

permutation has been read in the input string. We use '00' to refer to the dead

state in the second rank and '000' for the dead state in the third rank:



```

State(name=00, time=0)

State(name=000, time=0)

```



And this completes our encoding of the state machine into the boolean

satisfiability problem.



## An extension



As mentioned in the overview section, I did not manage to solve any unsolved

problems by this approach. The resulting SAT formula for N=6 has approximately

40 million clauses and having run a variety of SAT solvers for several days,

they did not manage to find a satisfying assignment for a length of 871, nor

prove that one is impossible.



However, before drawing a line under this approach, I decided to add an

extension in the hope this could help the SAT solvers work in a more incremental

fashion towards a solution. I tried to use the ideas from the algorithm

presented in [this blog post](http://www.njohnston.ca/2014/08/all-minimal-superpermutations-on-five-symbols-have-been-found/)

and model the same constraints in the SAT problem.



The basic idea is to limit the number of 'wasted symbols' that may occur at each

point in string. The hope is that this forces the SAT solver to eliminate

regions of the search space sooner than it would otherwise and more rapidly find

a solution or prove that one is impossible. I won't re-explain the high-level

approach of finding the "maximum number of permutation that fit into a string of

given length". Instead, I'll refer you to the blog post for more details about

how this works.



The following sections explain an extension of the reduction to incorporate

limiting the number of wasted symbols.



**1) Modelling a wasted symbol**



If the machine is in the dead state in its final rank then it must mean the

symbol that was just read is a wasted symbol. However, there's another case when

a symbol is wasted which is when a permutation has already occurred in the

string and the final state for that permutation is active again.



Handling this second case is problematic. In our approach we do the following:



a) Reverse the direction of implication for the Goal states:



```

State(name=321, time=5) -> Goal(name=321, time=5)

```



This has the effect of ensuring whenever a final state is reached, the goal

state for that permutation has its number set to the current time. However, it

also changes the semantics so that no final state can be reached more than once

since the number of the goal state must be unique.



At time of writing, it isn't known whether the minimal superpermutation for

N > 5 always contains one of each permutation. For 1 <= N <= 5, there is only

one of each permutation in the string, but it hasn't been proven whether this

property holds as N increases. However, for our purposes, this seems like a

reasonable restriction to make as an initial exploration of the problem.



b) Remove the ban on number zero for 'Goal' states



With the above change in place, each permutation must occur exactly once in the

string. However, we want our algorithm to run incrementally so we can discover

the maximum number of permutations that fit into strings of given length.

Therefore, we want to relax the requirement that each permutation must occur

exactly once and instead specify they can occur at most once.



This will allow us to control how many permutations must occur in the string.

How this is done is explained in more detail further on. For now, to change the

'exactly once' to an 'at most once', we remove the ban on the number for each

goal state being the number zero which we think of as another kind of 'dead'

state. This has the effect that goal states for permutations not in the string

can take on a value of zero which has the desired effect.



c) Introduce variables for the wasted symbols



Now that permutations may occur a maximum of once in string, The case when

'a permutation has already occurred in the string' can no longer happen. That

means we can reliably use the dead states in the machine's final rank to

indicate whether a symbol is wasted. For convenience, we introduce a new

variable for each point in time which is true if and only if the machine is in

the dead state at that time:



```

State(name=000, time=5) -> Wasted(time=5)

Wasted(time=5) -> State(name=000, time=5)

```



We now have a set of variables that are true if and only if wasted symbols occur

in the string at each point in time.



**2) Modelling the maximum number of wasted symbols**



We want to limit the number of wasted symbols that may occur at different points

in the string based on the known bounds for the maximum number of permutations

that fit into a string with some number of wasted symbols. To achieve this, we

need some means of encoding an 'at most k' constraint for the number of wasted

symbols and increasing the value of k at later points in the string.



There are several ways to encode an 'at most k' constraint. There is a useful

comparison in [this paper](http://www.cs.toronto.edu/~fbacchus/csc2512/at_most_k.pdf).



For our purposes, we choose the 'Sequential Counter Encoding'. The primary

reason for this is, after some analysis, we found a way to adapt this encoding

so the value of k can be varied at different points in the string. There's also

a natural way to extend this encoding to incorporate an 'at least k' constraint

which will be explained later.



Note: We think there's a small mistake in the paper in clause (5). The second

literal should be negated to specify the 'can't be an overflow' constraint.



Rather than re-explain how the encoding works, we'll instead focus on how the

value of k can be varied with reference to the paper. In short, we vary the

number of counter bits in each register so that it is equal to the desired value

of k at each point in time. We omit clauses (3) and (4) when there's no

corresponding counter bit in the previous register.



This is consistent with the encoding since we know the count represented by this

bit cannot occur in the previous register and therefore can by thought of as

being 'false'. Since false can imply true or false, these clauses can be

omitted as it does not constrain the current register's bit.



Finally, if the value of k is zero, i.e. no wasted symbols may occur at that

point in the string, we encode this by adding clauses contradicting the variable

for the wasted symbols.



```

-Wasted(time=0) ^

-Wasted(time=1) ^

-Wasted(time=2)

```



This is not handled in the paper because it only considers values 1 <= k <= n.



**3) Modelling minimum number of wasted symbols**



We also decided to encode constraints on the minimum number of wasted symbols

that must occur at points in the string. Since we know the maximum number of

permutations that fit into strings that waste some number of symbols, we also

know points in the string at which some number of wasted symbols must have

occurred. The hope, again, is this will help eliminate regions of the search

space and speed up solving the problem.



To encode a minimum on the number of wasted symbols, we modify our sequential

counter encoding once more. This time, we add a tautological clause for the

previous register's counter bit in place of clause (3) when there's no counter

bit in the current register because the corresponding number is less than the

minimum value of k:



```

Register(count=1, time=4) ^

Register(count=2, time=4) ^

Register(count=3, time=4)

```



This is consistent with the encoding since we know all counter bits below the

minimum value of k must have been set in the previous register, otherwise the

current register's count would be below the minimum. This also means we don't

need variables to represent counts below these minimum k values, reducing the

number of variables in our encoding.



**4) Determining the bounds**



For each position in the string, we want to specify a range of values for k

based on the ideas presented in

[the aforementioned blog post](http://www.njohnston.ca/2014/08/all-minimal-superpermutations-on-five-symbols-have-been-found/).

We do this by incrementally generating SAT problems of different length and

then running a SAT solver on them. If the problem can be satisfied it means it's

possible to fit the required number of permutations into the given length,

otherwise it is not and we can update our bounds with this new information.



This incremental approach means we're running the SAT solver on the same problem

several times with small variations in the clauses relating to the counter bits.

Many solvers provide an incremental API that can improve the performance in

cases like this called [IPASIR](http://www.cs.utexas.edu/users/moore/acl2/manuals/current/manual/index-seo.php/IPASIR____IPASIR).

The basic idea being that knowledge attained about the problem in one solve is

often reusable in a subsequent solve. However, we did not try this approach due

to the expected difficulty integrating this API with Rust but this might be an

interesting line of investigation.



## Closing remarks



As mentioned in the overview section, this approach was ultimately unsuccessful

in solving the more challenging superpermutation problems. For N=6, the solvers

don't get very far before they begin to take exponentially increasing amounts of

time to determine the satisfiability of each of the incremental subproblems.



We tried

[lingeling](http://fmv.jku.at/lingeling/),

[treengeling](http://fmv.jku.at/lingeling/),

[cadical](http://fmv.jku.at/cadical/),

[riss](http://tools.computational-logic.org/content/riss.php)

and [cryptominisat](https://www.msoos.org/cryptominisat5/) and observed similar

performance. Of these, treengeling is a parallel, cube-and-conquer solver which

we expected to perform the best because of reasons explained in

[this paper](https://www.cs.utexas.edu/~marijn/publications/CCshort.pdf) but

found lingeling and cadical perform better in practice.



If we had to speculate as to why the problem is particularly challenging for SAT

solvers, we think the following reasons are contributing factors:



- It's difficult to encode cardinality constraints in SAT which we use in our

extended approach. There are a lot of ways to do this and it's not clear how

each of the solvers will perform on each of them



- The superpermutation problem has similarities with the pidgeonhole problem in

that there are a lot of combinations to try and we're trying to fit some number

of permutations into strings of different length. For resolution-based SAT

solvers (of which most of them are), [it has been proven](https://arxiv.org/pdf/1705.01477.pdf)

that the pidgeonhole problem is a hard instance to solve



There's ongoing work to build SAT solvers that use richer proof systems, such

as [cutting planes](https://en.wikipedia.org/wiki/Cutting-plane_method), but for

the time being resolution- and CDCL-based solvers tend to outpace these solvers.



As I mentioned earlier, I'm sorry all the above isn't presented in a paper. I

think there's a lot of possible research directions that could be taken. In

particular, a formal analysis of various encodings and solver performance would

be an excellent line of investigation. Additionally, I think these problems are

particularly challenging for SAT solvers and this could help improve the solvers

themselves. I'd like to submit some of these problems as benchmarks in

[the annual SAT competitions](https://www.cs.helsinki.fi/u/mjarvisa/papers/jarvisalo-leberre-roussel-simon.aimag.pdf).



If you'd like to contact me about any of this work, my email address is listed

on [my GitHub profile](https://github.com/tuzz). Thanks.