Mini-kanren like logic programming in Scheme

sokuza-kanren.scm

;                    Quick miniKanren-like code 
;
; written at the meeting of a Functional Programming Group
; (Toukyou/Shibuya, Apr 29, 2006), as a quick illustration of logic
; programming.  The code is really quite trivial and unsophisticated:
; it was written without any preparation whatsoever. The present file
; adds comments and makes minor adjustments.
;
; $Id: sokuza-kanren.scm,v 1.1 2006/05/10 23:12:41 oleg Exp oleg $


; Point 1: `functions' that can have more (or less) than one result
;
; As known from logic, a binary relation xRy (where x \in X, y \in Y)
; can be represented by a _function_  X -> PowerSet{Y}. As usual in
; computer science, we interpret the set PowerSet{Y} as a multi-set
; (realized as a regular scheme list). Compare with SQL, which likewise
; uses multisets and sequences were sets are properly called for. 
; Also compare with Wadler's `representing failure as a list of successes.'
;
; Thus, we represent a 'relation' (aka `non-deterministic function')
; as a regular scheme function that returns a list of possible results.
; Here, we use a regular list rather than a lazy list, just to be quick.

; First, we define two primitive non-deterministic functions;
; one of them yields no result whatsoever for any argument; the other
; merely returns its argument as the sole result.

(define (fail x) '())
(define (succeed x) (list x))

; We build more complex non-deterministic functions by combining
; the existing ones with the help of the following two combinators.


; (disj f1 f2) returns all the results of f1 and all the results of f2.
; (disj f1 f2) returns no results only if neither f1 nor f2 returned
; any. In that sense, it is analogous to the logical disjunction.
(define (disj f1 f2)  
  (lambda (x)
    (append (f1 x) (f2 x))))

; (conj f1 f2) looks like a `functional composition' of f2 and f1.
; Only (f1 x) may return several results, so we have to apply f2 to
; each of them. 
; Obviously (conj fail f) and (conj f fail) are both equivalent to fail:
; they return no results, ever. It that sense, conj is analogous to the
; logical conjunction.
(define (conj f1 f2)
  (lambda (x)
    (apply append (map f2 (f1 x)))))


; Examples
(define (cout . args)
  (for-each display args))
(define nl #\newline)

(cout "test1" nl 
  ((disj 
     (disj fail succeed) 
     (conj 
       (disj (lambda (x) (succeed (+ x 1)))
	     (lambda (x) (succeed (+ x 10))))
       (disj succeed succeed)))
    100)
  nl)
; => (100 101 101 110 110)



; Point 2: (Prolog-like) Logic variables
;
; One may think of regular variables as `certain knowledge': they give
; names to definite values.  A logic variable then stands for
; `improvable ignorance'.  An unbound logic variable represents no
; knowledge at all; in other words, it represents the result of a
; measurement _before_ we have done the measurement. A logic variable
; may be associated with a definite value, like 10. That means
; definite knowledge.  A logic variable may be associated with a
; semi-definite value, like (list X) where X is an unbound
; variable. We know something about the original variable: it is
; associated with the list of one element.  We can't say though what
; that element is. A logic variable can be associated with another,
; unbound logic variable. In that case, we still don't know what
; precisely the original variable stands for. However, we can say that it
; represents the same thing as the other variable. So, our
; uncertainty is reduced.

; We chose to represent logic variables as vectors:
(define (var name) (vector name))
(define var? vector?)

; We implement associations of logic variables and their values 
; (aka, _substitutions_) as associative lists of (variable . value)
; pairs.
; One may say that a substitution represents our current knowledge
; of the world.

(define empty-subst '())
(define (ext-s var value s) (cons (cons var value) s))

; Find the value associated with var in substitution s.
; Return var itself if it is unbound.
; In miniKanren, this function is called 'walk'
(define (lookup var s)
  (cond
    ((not (var? var)) var)
    ((assq var s) => (lambda (b) (lookup (cdr b) s)))
    (else var)))

; There are actually two ways of implementing substitutions as
; associative list.
; If the variable x is associated with y and y is associated with 1,
; we could represent this knowledge as
; ((x . 1) (y . 1))
; It is easy to lookup the value associated with the variable then,
; via a simple assq. OTH, if we have the substitution ((x . y))
; and we wish to add the association of y to 1, we have
; to make rearrangements so to produce ((x . 1) (y . 1)).
; OTH, we can just record the associations as we learn them, without
; modifying the previous ones. If originally we knew ((x . y))
; and later we learned that y is associated with 1, we can simply
; prepend the latter association, obtaining ((y . 1) (x . y)).
; So, adding new knowledge becomes fast. The lookup procedure becomes
; more complex though, as we have to chase the chains of variables.
; To obtain the value associated with x in the latter substitution, we
; first lookup x, obtain y (another logic variable), then lookup y
; finally obtaining 1. 
; We prefer the latter, incremental way of representing knowledge:
; it is easier to backtrack if we later find out our
; knowledge leads to a contradiction.


; Unification is the process of improving knowledge: or, the process
; of measurement. That measurement may uncover a contradiction though
; (things are not what we thought them to be). To be precise, the
; unification is the statement that two terms are the same. For
; example, unification of 1 and 1 is successful -- 1 is indeed the
; same as 1. That doesn't add however to our knowledge of the world. If
; the logic variable X is associated with 1 in the current
; substitution, the unification of X with 2 yields a contradiction
; (the new measurement is not consistent with the previous
; measurements/hypotheses).  Unification of an unbound logic variable
; X and 1 improves our knowledge: the `measurement' found that X is
; actually 1.  We record that fact in the new substitution.


; return the new substitution, or #f on contradiction.
(define (unify t1 t2 s)
  (let ((t1 (lookup t1 s)) ; find out what t1 actually is given our knowledge s
	(t2 (lookup t2 s))); find out what t2 actually is given our knowledge s
    (cond
      ((eq? t1 t2) s)		; t1 and t2 are the same; no new knowledge
      ((var? t1)		; t1 is an unbound variable
	(ext-s t1 t2 s))
      ((var? t2)		; t2 is an unbound variable
	(ext-s t2 t1 s))
      ((and (pair? t1) (pair? t2)) ; if t1 is a pair, so must be t2
	(let ((s (unify (car t1) (car t2) s)))
	  (and s (unify (cdr t1) (cdr t2) s))))
      ((equal? t1 t2) s)	; t1 and t2 are really the same values
      (else #f))))


; define a bunch of logic variables, for convenience
(define vx (var 'x))
(define vy (var 'y))
(define vz (var 'z))
(define vq (var 'q))

(cout "test-u1" nl
  (unify vx vy empty-subst)
  nl)
; => ((#(x) . #(y)))

(cout "test-u2" nl
  (unify vx 1 (unify vx vy empty-subst))
  nl)
; => ((#(y) . 1) (#(x) . #(y)))

(cout "test-u3" nl
  (lookup vy (unify vx 1 (unify vx vy empty-subst)))
  nl)
; => 1 
; when two variables are associated with each other,
; improving our knowledge about one of them improves the knowledge of the
; other

(cout "test-u4" nl
  (unify (cons vx vy) (cons vy 1) empty-subst)
  nl)
; => ((#(y) . 1) (#(x) . #(y)))
; exactly the same substitution as in test-u2



; Part 3: Logic system
;
; Now we can combine non-deterministic functions (Part 1) and
; the representation of knowledge (Part 2) into a logic system.
; We introduce a 'goal' -- a non-deterministic function that takes
; a substitution and produces 0, 1 or more other substitutions (new
; knowledge). In case the goal produces 0 substitutions, we say that the
; goal failed. We will call any result produced by the goal an 'outcome'.

; The functions 'succeed' and 'fail' defined earlier are obviously
; goals.  The latter is the failing goal. OTH, 'succeed' is the
; trivial successful goal, a tautology that doesn't improve our
; knowledge of the world. We can now add another primitive goal, the
; result of a `measurement'.  The quantum-mechanical connotations of
; `the measurement' must be obvious by now.

(define (== t1 t2)
  (lambda (s)
    (cond
      ((unify t1 t2 s) => succeed)
      (else (fail s)))))


; We also need a way to 'run' a goal,
; to see what knowledge we can obtain starting from sheer ignorance
(define (run g) (g empty-subst))


; We can build more complex goals using lambda-abstractions and previously
; defined combinators, conj and disj.
; For example, we can define the function `choice' such that
; (choice t1 a-list) is a goal that succeeds if t1 is an element of a-list.

(define (choice var lst)
  (if (null? lst) fail
    (disj
      (== var (car lst))
      (choice var (cdr lst)))))

(cout "test choice 1" nl
  (run (choice 2 '(1 2 3)))
  nl)
; => (()) success

(cout "test choice 2" nl
  (run (choice 10 '(1 2 3)))
  nl)
; => ()
; empty list of outcomes: 10 is not a member of '(1 2 3)

(cout "test choice 3" nl
  (run (choice vx '(1 2 3)))
  nl)
; => (((#(x) . 1)) ((#(x) . 2)) ((#(x) . 3)))
; three outcomes

; The name `choice' should evoke The Axiom of Choice...

; Now we can write a very primitive program: find an element that is
; common in two lists:

(define (common-el l1 l2)
  (conj
    (choice vx l1)
    (choice vx l2)))

(cout "common-el-1" nl
  (run (common-el '(1 2 3) '(3 4 5)))
  nl)
; => (((#(x) . 3)))

(cout "common-el-2" nl
  (run (common-el '(1 2 3) '(3 4 1 7)))
  nl)
; => (((#(x) . 1)) ((#(x) . 3)))
; two elements are in common

(cout "common-el-3" nl
  (run (common-el '(11 2 3) '(13 4 1 7)))
  nl)
; => ()
; nothing in common


; Let us do something a bit more complex

(define (conso a b l) (== (cons a b) l))

; (conso a b l) is a goal that succeeds if in the current state
; of the world, (cons a b) is the same as l.
; That may, at first, sound like the meaning of cons. However, the
; declarative formulation is more powerful, because a, b, or l might
; be logic variables.
;
; By running the goal which includes logic variables we are
; essentially asking the question what the state of the world should
; be so that (cons a b) could be the same as l.

(cout "conso-1" nl
  (run (conso 1 '(2 3) vx))
  nl)
; => (((#(x) 1 2 3))) === (((#(x) . (1 2 3))))

(cout "conso-2" nl
  (run (conso vx vy (list 1 2 3)))
  nl)
; => (((#(y) 2 3) (#(x) . 1)))
; That looks now like 'cons' in reverse. The answer means that
; if we replace vx with 1 and vy with (2 3), then (cons vx vy)
; will be the same as '(1 2 3)

; Terminology: (conso vx vy '(1 2 3)) is a goal (or, to be more precise,
; an expression that evaluates to a goal). By itself, 'conso'
; is a parameterized goal (or, abstraction over a goal):
; conso === (lambda (x y z) (conso x y z))
; We will call such an abstraction 'relation'.

; Let us attempt a more complex relation: appendo
; That is, (appendo l1 l2 l3) holds if the list l3 is the
; concatenation of lists l1 and l2.
; The first attempt:

(define (apppendo l1 l2 l3)
  (disj
    (conj (== l1 '()) (== l2 l3))    ; [] ++ l == l
    (let ((h (var 'h)) (t (var 't))  ; (h:t) ++ l == h : (t ++ l)
	  (l3p (var 'l3p)))
      (conj
	(conso h t l1)
	(conj
	  (conso h l3p l3)
	  (apppendo t l2 l3p))))))

; If we run the following, we get into the infinite loop.
; (cout "t1"
;   (run (apppendo '(1) '(2) vq))
;   nl)

; It is instructive to analyze why. The reason is that 
; (apppendo t l2 l3p) is a function application in Scheme,
; and so the (call-by-value) evaluator tries to find its value first,
; before invoking (conso h t l1). But evaluating (apppendo t l2 l3p)
; will again require the evaluation of (apppendo t1 l21 l3p1), etc.
; So, we have to introduce eta-expansion. Now, the recursive
; call to apppendo gets evaluated only when conj applies
; (lambda (s) ((apppendo t l2 l3p) s)) to each result of (conso h l3p l3).
; If the latter yields '() (no results), then appendo will not be
; invoked. Compare that with the situation above, where appendo would 
; have been invoked anyway.

(define (apppendo l1 l2 l3)
  (disj                              ; In Haskell notation:
    (conj (== l1 '()) (== l2 l3))    ; [] ++ l == l
    (let ((h (var 'h)) (t (var 't))  ; (h:t) ++ l == h : (t ++ l)
	  (l3p (var 'l3p)))
      (conj
	(conso h t l1)
	(lambda (s)
	((conj
	  (conso h l3p l3)
	  (lambda (s)
	   ((apppendo t l2 l3p) s))) s))))))

(cout "t1" nl
  (run (apppendo '(1) '(2) vq))
  nl)
; => (((#(l3p) 2) (#(q) #(h) . #(l3p)) (#(t)) (#(h) . 1)))

; That all appears to work, but the result is kind of ugly;
; and all the eta-expansion spoils the code.

; To hide the eta-expansion (that is, (lambda (s) ...) forms),
; we have to introduce a bit of syntactic sugar:

(define-syntax conj*
  (syntax-rules ()
    ((conj*) succeed)
    ((conj* g) g)
    ((conj* g gs ...)
      (conj g (lambda (s) ((conj* gs ...) s))))))

; Incidentally, for disj* we can use a regular function
; (because we represent all the values yielded by a non-deterministic
; function as a regular list rather than a lazy list). All branches
; of disj will be evaluated anyway, in our present model.
(define (disj* . gs)
  (if (null? gs) fail
    (disj (car gs) (apply disj* (cdr gs)))))

; And so we can re-define appendo as follows. It does look
; quite declarative, as the statement of two equations that
; define what list concatenation is.

(define (apppendo l1 l2 l3)
  (disj                              ; In Haskell notation:
    (conj* (== l1 '()) (== l2 l3))   ; [] ++ l == l
    (let ((h (var 'h)) (t (var 't))  ; (h:t) ++ l == h : (t ++ l)
	  (l3p (var 'l3p)))
      (conj*
	(conso h t l1)
	(conso h l3p l3)
	(apppendo t l2 l3p)))))


; We also would like to make the result yielded by run more
; pleasant to look at.
; First of all, let us assume that the variable vq (if bound),
; holds the answer to our inquiry. Thus, our new run will try to
; find the value associated with vq in the final substitution.
; However, the found value may itself contain logic variables.
; We would like to replace them, too, with their associated values,
; if any, so the returned value will be more informative.

; We define a more diligent version of lookup, which replaces
; variables with their values even if those variables occur deep
; inside a term.

(define (lookup* var s)
  (let ((v (lookup var s)))
    (cond
      ((var? v) v)			; if lookup returned var, it is unbound
      ((pair? v)
	(cons (lookup* (car v) s)
	      (lookup* (cdr v) s)))
      (else v))))

; We can now redefine run as

(define (run g)
  (map (lambda (s) (lookup* vq s)) (g empty-subst)))

; and we can re-run the test

(cout "t1" nl
  (run (apppendo '(1) '(2) vq))
  nl)
; => ((1 2))

(cout "t2" nl
  (run (apppendo '(1) '(2) '(1)))
  nl)
; => ()
; That is, concatenation of '(1) and '(2) is not the same as '(1)

(cout "t3" nl
  (run (apppendo '(1 2 3) vq '(1 2 3 4 5)))
  nl)
; => ((4 5))


(cout "t4" nl
  (run (apppendo vq '(4 5) '(1 2 3 4 5)))
  nl)
; => ((1 2 3))

(cout "t5" nl
  (run (apppendo vq vx '(1 2 3 4 5)))
  nl)
; => (() (1) (1 2) (1 2 3) (1 2 3 4) (1 2 3 4 5))
; All prefixes of '(1 2 3 4 5)


(cout "t6" nl
  (run (apppendo vx vq '(1 2 3 4 5)))
  nl)
; => ((1 2 3 4 5) (2 3 4 5) (3 4 5) (4 5) (5) ())
; All suffixes of '(1 2 3 4 5)


(cout "t7" nl
  (run (let ((x (var 'x)) (y (var 'y)))
	 (conj* (apppendo x y '(1 2 3 4 5))
	        (== vq (list x y)))))
  nl)
; => ((() (1 2 3 4 5)) ((1) (2 3 4 5)) ((1 2) (3 4 5))
;     ((1 2 3) (4 5)) ((1 2 3 4) (5)) ((1 2 3 4 5) ()))
; All the ways to split (1 2 3 4 5) into two complementary parts


; For more detail, please see `The Reasoned Schemer'