A Primer¶
This short document aims to show the basic concepts underlying μkanren, namely relations as goals, and how to use primitive goals as building blocks to build and define relations. We start from two primitive relations toward recursive relations, through constructors, fresh logic variables, unification and η-inversion.
First of all we start importing objects and definitions of the logic system:
>>> from muk.core import *
>>> from muk.ext import *
>>> from muk.utils import *
Moreover, our definitions lies on the following imports:
>>> from collections import defaultdict, Counter
>>> from functools import lru_cache
We use the interface function run
to ask for
substitutions that satisfy the relation under study; moreover, nouns relation,
goal and predicate are equivalent notions and denotes the same concept in
the rest of this presentation.
Primitive goals¶
The two primitive goals succeed
and
fail
denote truth and falsehood in classic
logic, respectively.
succeed
¶
Asking for all substitutions that satisfy the succeed
relation yields a tautology, namely any substitution works:
>>> run(succeed)
[Tautology]
fail
¶
On the other hand, there is no substitution that satisfies the fail
relation:
>>> run(fail)
[]
Goal ctors¶
Before introducing goal constructors we show how a logic variable is represented when it doesn’t get any association:
>>> rvar(0)
▢₀
where class rvar
builds a reified variable providing an
integer, namely 0
, to track the variable’s birthdate.
fresh
¶
It is a goal ctor that introduces new logic variables, namely variables without
associations, according to the plugged-in lambda expression, which receives
vars and returns a goal. Consider the following query where the variable q
is introduced fresh:
>>> run(fresh(lambda q: succeed))
[▢₀]
since the returned goal is succeed
and var q
doesn’t get any
association, it remains fresh in the list of values, result of the whole
run
invocation, that satisfy the goal fresh(lambda q: succeed)
.
As particular case, it is possible to plugin a thunk, namely a lambda expression without argument:
>>> run(fresh(lambda: succeed))
[Tautology]
at a first this could look useless but it is of great help for the definition of recursive relations as we will see in later examples (it is an instance of η-inversion, formally).
unify
¶
It is a goal ctor that attempts to make two arbitrary objects equal, recording associations when fresh variables appears in the nested structures under unification. Here we show two simple examples of unification, the first succeeds while the second doesn’t:
>>> run(unify(3, 3))
[Tautology]
>>> run(unify([1, 2, 3], [[1]]))
[]
On the other hand, things get interesting when fresh variables are mixed in:
>>> run(fresh(lambda q: unify(3, q)))
[3]
>>> run(fresh(lambda q: unify([1, 2, 3], [1] + q)))
[[2, 3]]
>>> run(fresh(lambda q: unify([[2, 3], 1, 2, 3], [q, 1] + q)))
[[2, 3]]
When two fresh vars are unified it is said that they share or co-refer:
>>> run(fresh(lambda q, z: unify(q, z)))
[▢₀]
>>> run(fresh(lambda q, z: unify(q, z) & unify(z, 3)),
... var_selector=lambda q, z: q)
[3]
disj
¶
It is a goal ctor that consumes two goals and returns a new goal that can be satisfied when either the former or the latter goal can be satisfied:
>>> run(succeed | fail)
[Tautology]
>>> run(fail | fresh(lambda q: unify(q, True)))
[Tautology]
>>> run(fresh(lambda q: fail | fail))
[]
>>> run(fresh(lambda q: unify(q, False) | unify(q, True)))
[False, True]
conj
¶
It is a goal ctor that consumes two goals and returns a new goal that can be satisfied when both the former and the latter goal can be satisfied:
>>> run(succeed & fail)
[]
>>> run(fail & fresh(lambda q: unify(q, True)))
[]
>>> run(fresh(lambda q: unify(q, 3) & succeed))
[3]
>>> run(fresh(lambda q: unify(q, False) & unify(q, True)))
[]
>>> run(fresh(lambda q: fresh(lambda q: unify(q, False)) &
... unify(q, True)))
[True]
Facts and recursive relations¶
In order to represent facts we introduce the conde
goal ctor,
which is defined as a combination of conjs and disjs and we show how to write
recursive relation, possibly satisfied by a countably infinite number of values.
conde
¶
The following simple example resembles facts declaration in Prolog about the father and grandfather toy example:
>>> def father(p, s):
... return conde([unify(p, 'paul'), unify(s, 'jason')],
... [unify(p, 'john'), unify(s, 'henry')],
... [unify(p, 'jason'), unify(s, 'tom')],
... [unify(p, 'peter'), unify(s, 'brian')],
... [unify(p, 'tom'), unify(s, 'peter')])
...
>>> def grand_father(g, s):
... return fresh(lambda p: father(g, p) & father(p, s))
...
>>> run(fresh(lambda rel, p, s: grand_father(p, s) & unify([p, s], rel)))
[['paul', 'tom'], ['jason', 'peter'], ['tom', 'brian']]
Introducing guided search:
>>> def recuro(e):
... return fresh(lambda a, b: conde([unify(e, ['x']), succeed],
... [unify(e, ['b']+a), recuro(a)],
... [unify(e, ['a']+b), recuro(b)]))
...
>>> run(fresh(recuro), n=5)
[['x'], ['b', 'x'], ['b', 'b', 'x'], ['b', 'b', 'b', 'x'], ['b', 'b', 'b', 'b', 'x']]
η-inversion¶
Let us define a relation that yields countably many 5 objects; in order to do that, the usual solution is to write a recursive definition. However, we proceed step by step, adjusting and learning from the Python semantic of argument evaluation at function-call time. Consider the following as initial definition:
>>> def fives(x):
... return unify(5, x) | fives(x)
...
>>> run(fresh(lambda x: fives(x)))
Traceback (most recent call last):
...
RecursionError: maximum recursion depth exceeded while calling a Python object
Exception RecursionError
is raised because in the body of function fives
it
is required to evaluate fives(x)
in order to return a disj
object, but this is
the point from were we started, hence no progress for recursion.
Keeping in mind the previous argument, why not wrapping the recursion on
fives(x)
inside a fresh
ctor in order to refresh the var x
at
each invocation?
>>> def fives(x):
... return unify(5, x) | fresh(lambda x: fives(x))
...
>>> run(fresh(lambda x: fives(x)))
Traceback (most recent call last):
...
RecursionError: maximum recursion depth exceeded while calling a Python object
Again the same exception as before, this time for a different reason, however:
since we ask for all associations that satisfy the countably infinite
relation fives
, function run
continue to look for such values which are
infinite, of course. So, select only the first 10 objects:
>>> def fives(x):
... return unify(5, x) | fresh(lambda x: fives(x))
...
>>> run(fresh(lambda x: fives(x)), n=10)
[5, ▢₀, ▢₀, ▢₀, ▢₀, ▢₀, ▢₀, ▢₀, ▢₀, ▢₀]
Although not a list of 10 fives objects, it makes sense: the very first 5 gets
associated to the var x
introduced by the goal provided to run
by the
fresh
ctor, and this association is only one way to satisfy the disj
in
the definition of relation fives
. Looking for other associations that work,
we attempt to satisfy the second goal in the disj
, namely fresh(lambda x:
fives(x))
: it introduces a new var x
, different from the previous one,
and then recurs, leaving the original var without association. Since associations shown
in the output list refer to the very first var x
, we get many ▢₀
symbols
which represent the absence of association, therefore x
remains fresh.
fives
¶
One way to actually get a list of fives is to unify inside the inner fresh
, as follows:
>>> def fives(x):
... return unify(5, x) | fresh(lambda y: fives(y) & unify(y, x))
...
>>> run(fresh(lambda x: fives(x)), n=10)
[5, 5, 5, 5, 5, 5, 5, 5, 5, 5]
or to use fresh
as η-inversion rule, as follows:
>>> def fives(x):
... return unify(5, x) | fresh(lambda: fives(x))
...
>>> run(fresh(lambda x: fives(x)), n=10)
[5, 5, 5, 5, 5, 5, 5, 5, 5, 5]
nats
¶
Just for fun, using the previous trick and abstracting out the 5s, we can generate the naturals, taking only the first 10 as follows:
>>> def nats(x, n=0):
... return unify(n, x) | fresh(lambda: nats(x, n+1))
...
>>> run(fresh(lambda x: nats(x)), n=10)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
append
¶
Here we want to define the relation append(r, s, o)
which holds if
r + s == o
where r, s, o
are both lists. First of all we need an
helper relation nullo
which holds if l == []
:
>>> def nullo(l):
... return unify([], l)
so it follows the recursive definition, as usual:
>>> def append(r, s, out):
... def A(r, out):
... return conde([nullo(r), unify(s, out)],
... else_clause=fresh(lambda a, d, res:
... unify([a]+d, r) &
... unify([a]+res, out) &
... fresh(lambda: A(d, res))))
... return A(r, out)
Some examples follow:
>>> run(fresh(lambda q: append([1,2,3], [4,5,6], q)))
[[1, 2, 3, 4, 5, 6]]
>>> run(fresh(lambda l, q: append([1,2,3]+q, [4,5,6], l)), n=4)
[[1, 2, 3, 4, 5, 6],
[1, 2, 3, ▢₀, 4, 5, 6],
[1, 2, 3, ▢₀, ▢₁, 4, 5, 6],
[1, 2, 3, ▢₀, ▢₁, ▢₂, 4, 5, 6]]
>>> run(fresh(lambda r, x, y:
... append(x, y, ['cake', 'with', 'ice', 'd', 't']) &
... unify([x, y], r)))
[[[], ['cake', 'with', 'ice', 'd', 't']],
[['cake'], ['with', 'ice', 'd', 't']],
[['cake', 'with'], ['ice', 'd', 't']],
[['cake', 'with', 'ice'], ['d', 't']],
[['cake', 'with', 'ice', 'd'], ['t']],
[['cake', 'with', 'ice', 'd', 't'], []]]
Combinatorics¶
It is possible to generate the set of Dyck paths as it follows:
>>> def dycko(α):
... return conde([nullo(α), succeed],
... else_clause=fresh(lambda β, γ:
... append(['(']+β, [')']+γ, α) @ (dycko(β) @ dycko(γ))))
...
>>> paths = run(fresh(lambda α: dycko(α)), n=80)
>>> D = defaultdict(list)
>>> for α in map(lambda α: ''.join(α), paths):
... D[len(α)//2].append(α)
...
>>> dict(D)
{0: [''],
1: ['()'],
2: ['()()', '(())'],
3: ['()()()', '(())()', '()(())', '(()())', '((()))'],
4: ['(())()()', '()()()()', '(())(())', '()(())()', '(()()())', '(()())()', '()()(())', '(()(()))', '((()))()', '()(()())', '((())())', '()((()))', '((()()))', '(((())))'],
5: ['(())()()()', '(())(())()', '(()()()())', '(()()())()', '(()())()()', '(())()(())', '()(())()()', '(()()(()))', '((()))()()', '(())(()())', '()()()()()', '(()(()))()', '(()())(())', '((())()())', '((()))(())', '()(())(())', '(())((()))', '()()(())()', '((())())()', '()(()()())', '(()(())())', '()(()())()'],
6: ['(()()()()())', '(()()()())()', '(())(())()()', '(()()()(()))', '(()()())()()', '(())()()()()', '(()()(()))()', '(()())()()()', '((())()()())', '((()))()()()', '(())(())(())', '(()(()))()()', '(()())(())()', '(())()(())()', '((())()())()', '(()()())(())', '((()))(())()', '(())(()()())', '()(())()()()', '(()(())()())', '(()())()(())'],
7: ['(()()()()()())', '(()()()()())()', '(()()()()(()))', '(()()()())()()', '(()()()(()))()', '((())()()()())', '(()()(()))()()'],
8: ['(()()()()()()())', '(()()()()()())()', '(()()()()()(()))', '(()()()()())()()', '(()()()()(()))()'],
9: ['(()()()()()()()())', '(()()()()()()())()']}
Fibonacci numbers as tailing problem:
>>> def fibo(α):
... return conde([nullo(α), succeed],
... else_clause=fresh(lambda β: condi([unify(['v']+β, α), fibo(β)],
... [unify(['hh']+β, α), fibo(β)])))
...
>>> strips = run(fresh(lambda α: fibo(α)), n=80)
>>> D = defaultdict(list)
>>> for α in map(lambda α: ''.join(α), strips):
... D[len(α)].append(α)
...
>>> dict(D)
{0: [''],
1: ['v'],
2: ['hh', 'vv'],
3: ['hhv', 'vhh', 'vvv'],
4: ['hhhh', 'hhvv', 'vhhv', 'vvhh', 'vvvv'],
5: ['hhhhv', 'hhvhh', 'vhhhh', 'hhvvv', 'vhhvv', 'vvhhv', 'vvvhh', 'vvvvv'],
6: ['hhhhhh', 'hhhhvv', 'hhvhhv', 'vhhhhv', 'hhvvhh', 'vhhvhh', 'vvhhhh', 'hhvvvv', 'vhhvvv', 'vvhhvv', 'vvvhhv', 'vvvvhh', 'vvvvvv'],
7: ['hhhhhhv', 'hhhhvhh', 'hhvhhhh', 'vhhhhhh', 'hhhhvvv', 'hhvhhvv', 'vhhhhvv', 'hhvvhhv', 'vhhvhhv', 'vvhhhhv', 'hhvvvhh', 'vhhvvhh', 'vvhhvhh', 'vvvhhhh', 'hhvvvvv', 'vhhvvvv', 'vvhhvvv', 'vvvhhvv', 'vvvvhhv'],
8: ['hhhhhhhh', 'hhhhhhvv', 'hhhhvhhv', 'hhvhhhhv', 'vhhhhhhv', 'hhhhvvhh', 'hhvhhvhh', 'vhhhhvhh', 'hhvvhhhh', 'vhhvhhhh', 'vvhhhhhh', 'hhhhvvvv', 'hhvhhvvv', 'vhhhhvvv', 'hhvvhhvv', 'vhhvhhvv', 'vvhhhhvv'],
9: ['hhhhhhhhv', 'hhhhhhvhh', 'hhhhvhhhh', 'hhvhhhhhh', 'vhhhhhhhh', 'hhhhhhvvv', 'hhhhvhhvv', 'hhvhhhhvv', 'vhhhhhhvv'],
10: ['hhhhhhhhhh', 'hhhhhhhhvv']}
>>> run(fibo(['v', 'h', 'v']))
[]
Unfoldings within the Pascal triangle:
>>> from sympy import IndexedBase, symbols, latex
>>> P = IndexedBase('P')
>>> n, m = symbols('n m')
>>> def pascalo(depth, r, c, α):
... if not depth: return unify([P[r,c]], α)
... return fresh(lambda β, γ: (pascalo(depth-1, r-1, c-1, β) @
... pascalo(depth-1, r-1, c, γ) @
... append(β, γ, α)))
...
>>> unfoldings = [run(fresh(lambda α: pascalo(d, n, m, α))) for d in range(1, 6)]
>>> list(map(lambda unfold: sum(unfold[0]), unfoldings))
[P[n - 1, m - 1] + P[n - 1, m],
2*P[n - 2, m - 1] + P[n - 2, m - 2] + P[n - 2, m],
3*P[n - 3, m - 1] + 3*P[n - 3, m - 2] + P[n - 3, m - 3] + P[n - 3, m],
4*P[n - 4, m - 1] + 6*P[n - 4, m - 2] + 4*P[n - 4, m - 3] + P[n - 4, m - 4] + P[n - 4, m],
5*P[n - 5, m - 1] + 10*P[n - 5, m - 2] + 10*P[n - 5, m - 3] + 5*P[n - 5, m - 4] + P[n - 5, m - 5] + P[n - 5, m]]
Fibonacci rabbits, again:
>>> f = IndexedBase('f')
>>> def rabbitso(depth, r, α):
... if not depth: return unify([f[r]], α)
... return fresh(lambda β, γ: (rabbitso(depth-1, r-1, β) @
... rabbitso(depth-1, r-2, γ) @
... append(β, γ, α)))
...
>>> unfoldings = [run(fresh(lambda α: rabbitso(d, n, α))) for d in range(1, 6)]
>>> list(map(lambda unfold: sum(unfold[0]), unfoldings))
[f[n - 1] + f[n - 2],
f[n - 2] + 2*f[n - 3] + f[n - 4],
f[n - 3] + 3*f[n - 4] + 3*f[n - 5] + f[n - 6],
f[n - 4] + 4*f[n - 5] + 6*f[n - 6] + 4*f[n - 7] + f[n - 8],
f[n - 10] + f[n - 5] + 5*f[n - 6] + 10*f[n - 7] + 10*f[n - 8] + 5*f[n - 9]]
Interleaving¶
Let us now define an operator repeat
which makes relations that succeed by
unifying the curried variable with some plugged in object, infinitely many
times:
>>> def repeat(r):
... def R(x):
... return unify(x, r) | fresh(lambda: R(x))
... return R
...
consequently, use it to define four streams of numbers:
>>> ones = repeat(1)
>>> twos = repeat(2)
>>> threes = repeat(3)
>>> fours = repeat(4)
In the following example we use conde
, which yields solutions of a cond
line staying there when that line supplies multiple satisfying states, here
infinite when it meets fives
, by the way:
>>> def facts(x):
... return conde([unify(x, 'hello'), succeed],
... [unify(x, 'pluto'), succeed],
... [fives(x), succeed],
... [unify(x, 'topolino'), succeed],
... else_clause=unify(x, 'paperone'))
...
>>> run(fresh(lambda x: facts(x)), n=10)
['hello', 'pluto', 5, 5, 5, 5, 5, 5, 5, 5]
>>> groups_with_positions_notation(_)
hello₀
pluto₁
5₂, 5₃, 5₄, 5₅, 5₆, 5₇, 5₈, 5₉
On the other hand, condi
allows us to explore the solutions space by interleaving:
>>> def facts(x):
... return condi([unify(x, 'hello'), succeed],
... [unify(x, 'pluto'), succeed],
... [fives(x), succeed],
... [ones(x), succeed],
... else_clause=unify(x, 'paperone'))
...
>>> run(fresh(lambda x: facts(x)), n=15)
['hello', 'pluto', 5, 1, 5, 'paperone', 1, 5, 1, 5, 1, 5, 1, 5, 1]
>>> groups_with_positions_notation(_)
hello₀
pluto₁
5₂, 5₄, 5₇, 5₉, 5₁₁, 5₁₃
1₃, 1₆, 1₈, 1₁₀, 1₁₂, 1₁₄
paperone₅
Now, use only streams of numbers defined before and ask for the first 20
associations that satisfy their condi
combination, providing the keyword argument dovetail=False
which associates streams on the right:
>>> run(fresh(lambda q: condi([succeed, ones(q)],
... [succeed, twos(q)],
... [succeed, threes(q)],
... [succeed, fours(q)],
... dovetail=False)), n=20)
...
[1, 2, 1, 3, 1, 2, 1, 4, 1, 2, 1, 3, 1, 2, 1, 4, 1, 2, 1, 3]
>>> groups_with_positions_notation(_)
1₀, 1₂, 1₄, 1₆, 1₈, 1₁₀, 1₁₂, 1₁₄, 1₁₆, 1₁₈
2₁, 2₅, 2₉, 2₁₃, 2₁₇
3₃, 3₁₁, 3₁₉
4₇, 4₁₅
on the contrary, using dovetail=True
allows us to get a more balanced
interleaving (this is the default behaviour):
>>> run(fresh(lambda q: condi([succeed, ones(q)],
... [succeed, twos(q)],
... [succeed, threes(q)],
... [succeed, fours(q)])), n=20)
...
[1, 2, 1, 3, 2, 1, 4, 3, 2, 1, 4, 3, 2, 1, 4, 3, 2, 1, 4, 3]
>>> groups_with_positions_notation(_)
1₀, 1₂, 1₅, 1₉, 1₁₃, 1₁₇
2₁, 2₄, 2₈, 2₁₂, 2₁₆
3₃, 3₇, 3₁₁, 3₁₅, 3₁₉
4₆, 4₁₀, 4₁₄, 4₁₈
Previous expression is equivalent to combining streams with disj
getting
the same behavior of the last but one using the same keyword argument:
>>> run(fresh(lambda q: disj(ones(q), twos(q), threes(q), fours(q),
... dovetail=False)), n=20)
[1, 2, 1, 3, 1, 2, 1, 4, 1, 2, 1, 3, 1, 2, 1, 4, 1, 2, 1, 3]
>>> groups_with_positions_notation(_)
1₀, 1₂, 1₄, 1₆, 1₈, 1₁₀, 1₁₂, 1₁₄, 1₁₆, 1₁₈
2₁, 2₅, 2₉, 2₁₃, 2₁₇
3₃, 3₁₁, 3₁₉
4₇, 4₁₅
Finally, it is possible to request a fair enumeration by dovetail strategy,
setting the keyword argument dovetail=True
, which is the default behavior:
>>> run(fresh(lambda q: disj(ones(q), twos(q), threes(q), fours(q))), n=20)
[1, 2, 1, 3, 2, 1, 4, 3, 2, 1, 4, 3, 2, 1, 4, 3, 2, 1, 4, 3]
>>> groups_with_positions_notation(_)
1₀, 1₂, 1₅, 1₉, 1₁₃, 1₁₇
2₁, 2₄, 2₈, 2₁₂, 2₁₆
3₃, 3₇, 3₁₁, 3₁₅, 3₁₉
4₆, 4₁₀, 4₁₄, 4₁₈
As the last but one example, combining streams with the binary operator |
yields a yet different result because the disjunction binary operator
associates on the left:
>>> run(fresh(lambda q: ones(q) | twos(q) | threes(q) | fours(q)), n=20)
[1, 4, 3, 4, 2, 4, 3, 4, 1, 4, 3, 4, 2, 4, 3, 4, 1, 4, 3, 4]
>>> groups_with_positions_notation(_)
1₀, 1₈, 1₁₆
4₁, 4₃, 4₅, 4₇, 4₉, 4₁₁, 4₁₃, 4₁₅, 4₁₇, 4₁₉
3₂, 3₆, 3₁₀, 3₁₄, 3₁₈
2₄, 2₁₂
Difference structures¶
The following implementation of appendo
uses difference lists:
>>> def appendo(X, Y, XY):
... return fresh(lambda α, β, γ: unify(X, α-β) &
... unify(Y, β-γ) &
... unify(XY, α-γ))
...
and can be used as follows:
>>> run(fresh(lambda αβ, α, β: appendo(([1,2,3]+α)-α, ([4,5,6]+β)-β, αβ)))
[([1, 2, 3, 4, 5, 6] + ▢₀) - ▢₀]
>>> run(fresh(lambda out, X, Y, α, β:
... appendo(([1,2,3]+α)-α, ([4,5,6]+β)-β, X-Y) &
... unify([X, Y], out)))
[[[1, 2, 3, 4, 5, 6] + ▢₀, ▢₀]]