π΅οΈββοΈ Holmes
Holmes is a library for computing constraint-solving problems. Under the hood, it uses propagator networks and conflict-directed clause learning to optimise the search over the parameter space.
Now available on Hackage!
π Example
Dinesman's
problem
is a nice first example of a constraint problem. In this problem, we imagine
five people β Baker, Cooper, Fletcher, Miller, and Smith β living in a
five-story apartment block, and we must figure out the floor on which each
person lives. Here's how we state the problem with Holmes
:
import Data.Holmes
dinesman :: IO (Maybe [ Defined Int ])
dinesman = do
let guesses = 5 `from` [1 .. 5]
guesses `satisfying` \[ baker, cooper, fletcher, miller, smith ] -> and'
[ distinct [ baker, cooper, fletcher, miller, smith ]
, baker ./= 5
, cooper ./= 1
, fletcher ./= 1 .&& fletcher ./= 5
, miller .> cooper
, abs' (smith .- fletcher) ./= 1
, abs' (fletcher .- cooper) ./= 1
]
π£ Step-by-step problem-solving
Now we've written the poster example, how do we go about stating and solving our own constraint problems?
βοΈ 0. Pick a parameter type
Right now, there are two parameter type constructors: Defined
and
Intersect
. The choice of type determines the strategy by which we solve
the problem:
-
Defined
only permits two levels of knowledge about a value: nothing and everything. In other words, it doesn't support a notion of partial information; we either know a value, or we don't. This is fine for small problem spaces, particularly when few branches are likely to fail, but we can usually achieve faster results using another type. -
Intersect
stores a set of "possible answers", and attempts to eliminate possibilities as the computation progresses. For problems with many constraints, this will produce significantly faster results thanDefined
as we can hopefully discover failures much earlier.
It would seem that Intersect
would be the best choice in most cases, but
beware: it will only work for small enum types. An Intersect Int
for
which we have no knowledge will contain every possible Int
, and will
therefore take an intractable time to compute. Defined
has no such
restrictions.
πΊ 1. State the parameter space
Next, we need to produce a Config
stating the search space we want to explore
when looking for satisfactory inputs. The simplest way to do this is with the
from
function:
from :: Int -> [ x ] -> Config Holmes (Defined x)
from :: Int -> [ x ] -> Config Holmes (Intersect x)
If, for example, we wanted to solve a Sudoku problem, we might say something like this:
definedConfig :: Config Holmes (Defined Int)
definedConfig = 81 `from` [ 1 .. 9 ]
We read this as, "81
variables whose values must all be numbers between 1
and 9
". At this point, we place no constraints (such as uniqueness of rows or
columns); we're just stating the possible range of values that could exist in
each parameter.
We could do the same for Intersect
, but we'd first need to produce some
enum type to represent our parameter space:
data Value = V1 | V2 | V3 | V4 | V5 | V6 | V7 | V8 | V9
deriving stock (Eq, Ord, Show, Enum, Bounded, Generic)
deriving anyclass (Hashable)
instance Num Value where -- Syntactic sugar for numeric literals.
fromInteger = toEnum . pred . fromInteger
Now, we can produce an Intersect
parameter space. Because we can now work
with a type who has only 9
values, rather than all possible Int
values,
producing the initial possibilities list becomes tractable:
intersectConfig :: Config Holmes (Intersect Value)
intersectConfig = 81 `from` [ 1 .. 9 ]
There's one more function that lets us do slightly better with an Intersect
strategy, and that's using
:
using :: [ Intersect Value ] -> Config Holmes (Intersect Value)
With using
, we can give a precise "initial state" for all the Intersect
variables in our system. This, it turns out, is very convenient when we're
trying to state sudoku problems:
squares :: Config Holmes (Intersect Value)
squares = let x = mempty in using
[ x, 5, 6, x, x, 3, x, x, x
, 8, 1, x, x, x, x, x, x, x
, x, x, x, 5, 4, x, x, x, x
, x, x, 4, x, x, x, x, 8, 2
, 6, x, 8, 2, x, 4, 3, x, 7
, 7, 2, x, x, x, x, 4, x, x
, x, x, x, x, 7, 8, x, x, x
, x, x, x, x, x, x, x, 9, 3
, x, x, x, 3, x, x, 8, 2, x
]
Now, let's write some constraints!
π― 2. Declare your constraints
Typically, your constraints should be stated as a predicate on the input
parameters, with a type that, when specialised to your problem, should look
something like [Prop Holmes (Defined Int)] -> Prop Holmes (Defined Bool)
.
Now, what's this Prop
type?
πΈ Propagators
If this library has done its job properly, this predicate shouldn't look too
dissimilar to regular predicates. However, behind the scenes, the Prop
type
is wiring up a lot of relationships.
As an example, consider the (+)
function. This has two inputs and one output,
and the output is the sum of the two inputs. This is totally fixed, and there's
nothing we can do about it. This is fine when we write normal programs, because
we only have one-way information flow: input flows to output, and it's as
simple as that.
When we come to constraint problems, however, we have multi-way information flow: we might know the output before we know the inputs! Ideally, it'd be nice in these situations if we could "work backwards" to the information we're missing.
When we say x .+ y .== z
, we actually wire up multiple relationships:
x + y = z
, z - y = x
, and z - x = y
. That way, as soon as we learn
two of the three values involved in addition, we can infer the other!
The operators provided by this library aim to maximise information flow
around a propagator network by automatically wiring up all the different
identities for all the different operators. We'll see later that this
allows us to write seemingly-magical functions like backwards
: given a
function and an output, we can produce the function's input!
π The problem-solving toolkit
With all this in mind, the following functions are available to us for multi-directional information flow. We'll leave the type signatures to Haddock, and instead just run through the functions and either their analogous regular functions or a brief explanation of what they do:
π Boolean functions
Function | Analogous function / notes |
---|---|
(.&&) |
(&&) |
all' |
all |
allWithIndex' |
all' , but the predicate also receives the list index |
and' |
and |
(.||) |
(||) |
any' |
any |
anyWithIndex' |
any' , but the predicate also receives the list index |
or' |
or |
not' |
not |
false |
False |
true |
True |
π³οΈβπ Equality functions
Function | Analogous function / notes |
---|---|
(.==) |
(==) |
(./=) |
(/=) |
distinct |
Are all list elements different (according to (./=) )? |
π₯ Comparison functions
Function | Analogous function / notes |
---|---|
(.<) |
(<) |
(.<=) |
(<=) |
(.>) |
(>) |
(.>=) |
(>=) |
π Arithmetic functions
Function | Analogous function / notes |
---|---|
(.*) |
(*) |
(./) |
(/) |
(.+) |
(+) |
(.-) |
(-) |
(.%.) |
mod |
(.*.) |
(*) for integral functions |
(./.) |
div |
abs' |
abs |
negate' |
negate |
recip' |
recip |
π± Information-generating functions
Function | Analogous function / notes |
---|---|
(.$) |
Apply a function to the value within the parameter type. |
zipWith' |
Similar to liftA2 ; generate results from the parameters. |
(.>>=) |
Turn each value within the parameter type into the parameter type. |
The analogy gets stretched a bit here, unfortunately. It's perhaps helpful to
think of these functions in terms of Intersect
:
-
(.$)
maps over the remaining candidates in anIntersect
. -
zipWith'
creates anIntersect
of the cartesian product of the two givenIntersect
s, with the pairs applied to the given function. -
(.>>=)
takes every remaining candidate, applies the given function, then unions the results to produce anIntersect
of all possible results.
Using the above toolkit, we could express the constraints of our sudoku
example. After we establish some less interesting functions for splitting up
our 81
inputs into helpful chunks...
rows :: [ x ] -> [[ x ]]
rows [] = []
rows xs = take 9 xs : rows (drop 9 xs)
columns :: [ x ] -> [[ x ]]
columns = transpose . rows
subsquares :: [ x ] -> [[ x ]]
subsquares xs = do
x <- [ 0 .. 2 ]
y <- [ 0 .. 2 ]
let subrows = take 3 (drop (y * 3) (rows xs))
values = foldMap (take 3 . drop (x * 3)) subrows
pure values
... we can use the propagator toolkit to specify our constraints in a delightfully straightforward way:
constraints :: forall m. MonadCell m => [ Prop m (Intersect Value) ] -> Prop m (Intersect Bool)
constraints board = and'
[ all' distinct (columns board)
, all' distinct (rows board)
, all' distinct (subsquares board)
]
The type signature looks a little bit ugly here, but the polymorphism is to guarantee that predicate computations are totally generic propagator networks that can be run in any interpretation strategy. As we'll see later,
Holmes
isn't the only one capable of solving a mystery!Typically, we write the constraint predicate inline (as we did for the Dinesman example above), so we never usually write this signature anyway!)
We've explained all the rules and constraints of the sudoku puzzle, and designed a propagator network to solve it! Now, why don't we get ourselves a solution?
π‘ 3. Solving the puzzle
Currently, Holmes
only exposes two strategies for solving constraint
problems:
-
satisfying
, which returns the first valid configuration that is found, if one exists. As soon as this result has been found, computation will cease, and this program will return the result. -
whenever
, which returns all valid configurations in the search space. This function could potentially run for a long time, depending on the size of the search space, so you might find better results by sticking tosatisfying
and simply adding more constraints to eliminate the results you don't want!
These functions are named to be written as infix functions, which hopefully makes our programs a lot easier to read:
sudoku :: IO (Maybe [ Intersect Value ])
sudoku = squares `satisfying` constraints
At last, we combine the three steps to solve our problem. This README is a
literate Haskell file containing a complete sudoku solver, so feel free
to run cabal new-test readme
and see for yourself!
π Bonus surprises
We've now covered almost the full API of the library. However, there are a couple extra little surprises in there for the curious few:
π Control.Monad.Watson
Watson
knows Holmes
' methods, and can apply them to compute results. Unlike
Holmes
, however, Watson
is built on top of ST
rather than IO
, and is
thus a much purer soul.
Users can import Control.Monad.Watson
and use the equivalent satisfying
and
whenever
functions to return results without the IO
wrapper, thus making
these computations observably pure! For most computations β certainly those
outlined in this README β Watson
is more than capable of deducing results.
π² Random restart with shuffle
Watson
isn't quite as capable as Holmes
, however. Consider a typical
Config
:
example :: Config Holmes (Defined Int)
example = 1 `from` [1 .. 10]
With this Config
, a program will run with a single parameter. For the first
run, that parameter will be set to Exactly 1
. For the second run, it will
be set to Exactly 2
. In other words, it tries each value in order.
For many problems, however, we can get to results faster β or produce more
desirable results β by applying some randomness to this order. This is
especially useful in problems such as procedural generation, where
randomness tends to lead to more natural-seeming outputs. See the
WaveFunctionCollapse
example for more details!
β»οΈ Running functions forwards and backwards
With satisfying
and whenever
, we build a predicate on the input
parameters we supply. However, we can use propagators to create normal
functions, too! Consider the following function:
celsius2fahrenheit :: MonadCell m => Prop m (Defined Double) -> Prop m (Defined Double)
celsius2fahrenheit c = c .* (9 ./ 5) .+ 32
This function converts a temperature written in celsius to fahrenheit. The interesting part of this, however, is that this is a function over propagator networks. This means that, while we can use it as a regular function...
fahrenheit :: Maybe (Defined Double)
fahrenheit = forward celsius2fahrenheit 40.0 -- Just 104.0
... the "input" and "output" labels are meaningless! In fact, we can just as easily pass a value to the function as the output and get back the input!
celsius :: Maybe (Defined Double)
celsius = backward celsius2fahrenheit 104.0 -- Just 40.0
Because neither
forward
norbackward
require any parameter search, they are both computed byWatson
, so the results are pure!
π Exploring the code
Now we've covered the what, maybe you're interested in the how! If you're new to the code and want to get a feel for how the library works:
-
The best place to start is probably in
Data/JoinSemilattice/Class/*
(we can ignoreMerge
until the next step). These will give you an idea of how we represent relationships (as opposed to functions) inHolmes
. -
After that,
Control/Monad/Cell/Class.hs
gives an overview of the primitives for building a propagator network. In particular, seeunary
andbinary
for an idea of how we lift our relationships into a network. Here's wheresrc/Data/JoinSemilattice/Class/Merge
gets used, too, so thewrite
primitive should give you an idea of why it's useful. -
src/Data/Propagator.hs
introduces the high-level user-facing abstraction for stating constraints. Most of these functions are just wrapped calls to the aforementionedunary
orbinary
, and really just add some syntactic sugar. -
Finally,
Control/Monad/MoriarT.hs
is a full implementation of the interface including support for provenance and backtracking. It also uses the functions inData/CDCL.hs
to optimise the parameter search. This is the base transformer on top of which we buildControl/Monad/Holmes.hs
andControl/Monad/Watson.hs
.
Thus concludes our whistle-stop tour of my favourite sights in the repository!
βοΈ Questions?
If anything isn't clear, feel free to open an issue, or just message me on Twitter; it's where you'll most likely get a reply! I want this project to be an accessible way to approach the fields of propagators, constraint-solving, and CDCL. If there's anything I can do to improve this repository towards that goal, please let me know!
π Acknowledgements
-
Edward Kmett, whose propagators repository* gave us the
Prop
abstraction. I spent several months looking for alternative ways to represent computations, and never came close to something as neat. -
Marco Sampellegrini, Alex Peitsinis, Irene Papakonstantinou, and plenty others who have helped me figure out how to present this library in a maximally-accessible way.
* This repository also approaches propagator network computations using Andy
Gill's observable sharing
methods, which may be of interest! Neither Holmes
nor Watson
implement
this, as it requires some small breaks to purity and referential transparency,
of which users must be aware. We sacrifice some performance gains for ease of
use.