blog/posts/digitle-in-maude.md

13 KiB
Raw Blame History

mod DIGITLE is
  protecting NAT .
  vars M N : Nat .

the file starts with a module declaration. everything has to live inside some module, but this program isn't big enough to be worth splitting up so it's all going into one. since digitle involves numbers, i'm also importing the NAT module, and saying that any time M or N show up in equations or rewrite rules they stand for single numbers.

:::aside why is importing called protecting? in maude, sorts (approximately, types) aren't necessarily closed. you can add a new constructor, or equations between constructor forms, at any time. unlike in languages like haskell or rust, constructors don't have to be free, which means you can have equations between them. like to get /4 you can just do something like

fmod NAT4 is
  including NAT .
  eq 4 = 0 .
endfm

but when you're importing a module, most of the time you don't want to do any of that. so there are three ways of importing a module to express how much you intend to mess with it. protecting says you are not adding new constructors or conflating existing ones. extending lets you add new constructors, but not new equations. including lets you do whatever. since i'm just using normal numbers, protecting is what i want.

an important caveat is that it is not possible for maude itself to check you are telling the truth here, but it can ask yices or cvc4 for help. please don't press me for more detail i have never used this :::

the first thing i need is a type sort for the pool of numbers we currently have. it's a multiset: the number of copies of each number matters, but the order they're in doesn't. the most convenient way to express that is like this.

sort Pool .
subsort Nat < Pool .
op nil : -> Pool .
op __  : Pool Pool -> Pool [assoc comm  id: nil] .
var Ps : Pool .

first i just declare Pool as a sort, without saying anything about what it looks like. next i say that any Nat is also a Pool (containing one copy of just that number), by making Nat a subsort of Pool.

in the third and fourth lines, i say that nil is a pool (with no numbers), and two Pools written next to each other are also a Pool. so far, we've defined a binary tree. by making __ associative, we're saying that the exact bracketing of the appends doesn't matter, which flattens it to a list. by saying it's commutative, we're saying the order also doesn't matter.

the last attribute, id: nil, says that Ps nil = nil Ps = Ps. it also allows a pattern like M N Ps to match against 1 2, setting Ps = nil.

it feels a bit weird at times to be specifying constructors of a tree and then telling the language to flatten them after the fact, coming from languages where you try to design your datatypes to only have one representation per value. but the benefit of writing it like this instead of something like op cons : Nat Pool -> Pool is that maude knows what assoc, comm, and id mean, so later when i have some equations and rewrite rules that match on the beginning of a list, it knows that it can actually pick those numbers from anywhere, without needing extra rules to shuffle the list around manually.

:::aside the name __ is pretty weird! generally operator names are written with an underscore where the arguments go, so the name of the addition operator is _+_. in this case, there is no operator, it's just two things next to each other (in a context where a Pool value is expected, anyway), so it's just two underscores. :::

from here, we could just define the rewrite rules and be done. something like this would work:

rl  [add] : M N  =>  M + N .
rl  [sub] : M N  =>  sd(M, N) .
rl  [mul] : M N  =>  M * N .
crl [div] : M N  =>  M quo N if M rem N = 0 .

:::aside sd is "symmetric difference"---e.g. sd(1, 4) = sd(4, 1) = 3.
(it is also a declaration form which is why it's purple. i don't know how to distinguish the two in pandoc/kate's syntax highlighting format sorry)

:::

the first three rules just pick two numbers from our pool (any two, because of comm; they don't have to be next to each other), and replace them with the result of applying them to one of our operations. the last one is slightly more complicated, because we can only divide evenly. _quo_ ignores the remainder altogether, so we need a conditional rule which only fires if the remainder actually is zero.

save what we have so far, plus a terminating endm for the mod, in a file digitle.maude, and feed it into the repl:

$ maude digitle
                     \||||||||||||||||||/
                   --- Welcome to Maude ---                   
                     /||||||||||||||||||\
            Maude 3.2.1 built: Mar 13 2022 18:56:15
             Copyright 1997-2022 SRI International
                   Mon Mar 14 02:28:22 2022

let's use random puzzle ALO7b9UK as an example. we have to get to 793 starting from 75, 4, 7, 9, 8, 2. so starting from that list, we want to search for a sequence of rewrites that leads to a pool containing 793. one solution is all we need.

Maude> search [1] (75 4 7 9 8 2) =>* (793 Ps) .
search [1] in DIGITLE : 75 4 7 9 2 8 =>* Ps 793 .

Solution 1 (state 9577)
states: 9578  rewrites: 273357 in 128ms cpu (128ms real) (2135601 rewrites/second)
Ps --> 7

ok, so there is at least one solution, which leaves 7 unused (because it is in the assignment for Ps). we can use the state label 9577 to replay the sequence that reaches it, and see which steps it took.

Maude> show path 9577 .
state 0, Pool: 2 4 7 8 9 75
===[ rl N M Ps => Ps N + M [label add] . ]===>
state 4, Pool: 4 7 8 11 75
===[ rl N M Ps => Ps N * M [label mul] . ]===>
state 146, Pool: 7 11 32 75
===[ rl N M Ps => Ps N * M [label mul] . ]===>
state 1733, Pool: 7 32 825
===[ rl N M Ps => Ps sd(N, M) [label sub] . ]===>
state 9577, Pool: 7 793

squinting at the available numbers each time, along with the rule labels, we can just about make out what it did to get to a solution:

:::twocol

  • 2 + 9 = 11
  • 4 × 8 = 32
  • 11 × 75 = 825
  • 825 32 = 793 :::

ok that's cool. but reading the output is a bit annoying. let's instead keep track of what we did so that it's just printed legibly at the end of search. thanks to maude's flexible expression syntax, we can make this look like pretty much whatever we want.

sort Op .
ops + - × ÷ : -> Op .

sort Steps .

--- empty list
op nil : -> Steps .

--- a single step, like "3 + 4 → 7"
op ___→_ : Nat Op Nat Nat -> Steps [prec 10] .

--- sequence of steps
op _,_ : Steps Steps -> Steps [assoc id: nil prec 20] .

var Ss : Steps .

:::aside prec N is parsing/printing precedence (higher is looser, same as coq, opposite of haskell/agda/idris). :::

so solution traces are written like 2 + 9 → 11,4 × 8 → 32,11 × 75 → 825,825 - 32 → 793 and that is also how maude will print them.

the given state at any point is now going to be the available pool of numbers, plus the steps taken so far. this is just a pair, along with an abbreviation \{Ps} to make the search command look a little nicer.

sort State .
op _&_ : Pool Steps -> State .

op {_} : Pool -> State .
eq {Ps} = Ps & nil .

var S : State .

one last thing before the expanded rewrite rules is a predicate to say what we are looking for. this isn't strictly necessary, you could just continue to pattern match in the search command like last time; i just think this looks a bit nicer. but a state "has" a number N if N occurs anywhere in the pool. maude supports repeating pattern variables so this is nice and short.

op _has_ : State Nat -> Bool .
eq (N Ps & Ss) has N = true .
eq S           has N = false [otherwise] .

:::aside it could even skip the false line with a slightly different signature but then i would have to explain more about the sort & kind system... this is good enough. :::

the new rewrite rules (delete the other ones) have the same behaviour for the number pool, and also append the current step to the trace. since we don't need show path any more i removed the labels.

rl  M N Ps & Ss  =>  Ps (M + N)   & Ss, (M + N → (M + N)) .
rl  M N Ps & Ss  =>  Ps sd(M, N)  & Ss, (M - N → sd(M,N)) .
rl  M N Ps & Ss  =>  Ps (M * N)   & Ss, (M × N → (M * N)) .
crl M N Ps & Ss  =>  Ps (M quo N) & Ss, (M ÷ N → (M quo N)) if M rem N = 0 .

now load the updated file into maude and run the search command. i'm using s.t. (short for such that) to identify solutions by a boolean expression, the _has_ function above.

$ maude -no-banner digitle
Maude> search [1] {75 4 7 9 8 2} =>* S s.t. S has 793 .
search [1] in DIGITLE : {75 4 7 9 2 8} =>* S such that S has 793 = true .

Solution 1 (state 335734)
states: 335735  rewrites: 1114984 in 832ms cpu (830ms real) (1340125 rewrites/second)
S --> (7 793) & (9 + 2 → 11),(8 × 4 → 32),(75 × 11 → 825),(825 - 32 → 793)

and now we have our solution nicely printed!

there is one small thing that is still bothering me. if we are trying M + N, then trying N + M later is just wasting time. so what if we make all the rules conditional so that the first argument is never smaller than the second.

crl M N Ps & Ss  =>  Ps (M + N)   & Ss, (M + N → (M + N)) if M >= N .
crl M N Ps & Ss  =>  Ps sd(M, N)  & Ss, (M - N → sd(M,N)) if M >  N .
crl M N Ps & Ss  =>  Ps (M * N)   & Ss, (M × N → (M * N)) if M >= N .
crl M N Ps & Ss  =>  Ps (M quo N) & Ss, (M ÷ N → (M quo N))
  if M >= N /\ M rem N = 0 .

and look, it's five times faster now it's not repeating itself.

Maude> search [1] {75 4 7 9 8 2} =>* S s.t. S has 793 .
search [1] in DIGITLE : {75 4 7 9 2 8} =>* S such that S has 793 = true .

Solution 1 (state 46714)
states: 46715  rewrites: 270529 in 160ms cpu (158ms real) (1690806 rewrites/second)
S --> (7 793) & (9 + 2 → 11),(8 × 4 → 32),(75 × 11 → 825),(825 - 32 → 793)

here's the full file. the format directives i didn't discuss: the one in _,_ puts a space after the comma, and the rest do some colours & indenting to make the solution a little bit prettier. it also contains a predicate _has!_ for Hard Mode, that requires all the numbers get used up; and _almost has_, which finds any almost-solution within 10 of the target, and uses language features that maybe i will discuss if i ever write a second post about maude.

mod DIGITLE is
  protecting NAT .
  var M N : Nat .

  sort Pool .
  subsort Nat < Pool .
  op nil : -> Pool .
  op __  : Pool Pool -> Pool [assoc comm  id: nil] .
  var Ps : Pool .

  sort Op .
  op + : -> Op [format (r o)] .
  op - : -> Op [format (g o)] .
  op × : -> Op [format (y o)] .
  op ÷ : -> Op [format (b o)] .

  sort Steps .
  op nil : -> Steps .
  op ___→_ : Nat Op Nat Nat -> Steps [prec 20  format (d d d d ! o)] .
  op _,_ : Steps Steps -> Steps [assoc  id: nil  prec 10  format (d d s d)] .
  var Ss : Steps .

  sort State .
  op _&_ : Pool Steps -> State [format (d d n++i --)] .
  op {_} : Pool -> State .
  eq {Ps} = Ps & nil .
  var S : State .

  op _has_ : State Nat -> Bool .
  eq (N Ps & Ss) has N = true .
  eq S           has N = false [otherwise] .

  op _has!_ : State Nat -> Bool .
  eq (N & Ss) has! N = true .
  eq S        has! N = false [otherwise] .

  op _almost has_ : State Nat -> [Bool] .
  ceq (M Ps & Ss) almost has N = true if sd(M,N) <= 10 .

  crl M N Ps & Ss  =>  Ps (M + N)   & Ss, (M + N → (M + N)) if M >= N .
  crl M N Ps & Ss  =>  Ps sd(M, N)  & Ss, (M - N → sd(M,N)) if M >  N .
  crl M N Ps & Ss  =>  Ps (M * N)   & Ss, (M × N → (M * N)) if M >= N .
  crl M N Ps & Ss  =>  Ps (M quo N) & Ss, (M ÷ N → (M quo N))
    if M >= N /\ M rem N = 0 .
endm