-
Notifications
You must be signed in to change notification settings - Fork 0
Choices
One of the key data and control structures of Kno is the choice which represents a set of alternative values to be returned or explored by a function.
Operations over choices allow the natural expression of many search, set, and iteration operations. Choices are based on John McCarthy's AMB operator [original paper] where execution automatically explores multiple compute paths.
The printed representation of a choice consists of the alternatives separated by spaces and surrounded by curly braces ({}). Here's an example of a choice in Scheme:
{"alpha" "beta" (gamma delta)}
which is a choice of three elements: two strings and a list consisting of two symbols. The elements of choices can be any object except for another choice.
When procedures are applied to choices, they automatically iterate over the alternatives in their arguments and returns the alternative results as a choice. With multiple arguments, the procedure iterates over all the combinations of the values in the passed arguments (the cross product).
(list {1 2 3} {"one" "two"}
{(1 "one")
(2 "one")
(3 "one")
(1 "two")
(2 "two")
(3 "three")}
Like a mathematical set, choices are unordered collections of values without duplicates. Unlike a set, a choice can't be an element in a choice. In Kno, normal data values are treated as choices with one alternative.
Compound structures (pairs, vectors, tables, etc) can have choices as elements. As a knowledge representation language, choices are used to represent multi-valued slots. As a database foundation, choices are used to represent mappings of keys to the objects they characterize.
Certain procedures are declared to be non-deterministic meaning
that they pass their choice arguments directly without iterating over
them. In Scheme, you can create a non-deterministic procedure using
the special forms AMBDA
instead of LAMBDA
and DEFAMBDA
instead
of DEFINE
.
The C implementation of choices is optimized for performance. Choices
are immutable objects which are sorted to provide O(n)
performance
on set operations and O(log n)
performance on membership tests.
In the underlying C library, a special PRECHOICE
object holds
collection of objects which have not yet been sorted and reduced into
a choice. This significantly improves performance in many cases.
Choices are distinct from the multiple values provided by Common Lisp and the R5RS Scheme standard. Mutiple value implementations allow a procedure to return structured multiple values, where different value positions have different semantics (e.g. the first value might be an x coordinate and the second value might be a y coordinate). Choices in Kno, on the other hand, represent unstructured sets of values.
Curly braces represent literal choices, so evaluating a choice between 3, 4, and 5 just returns a choice between the three numbers.
#|kno>|# {3 4 5}
{3 4 5}
However, adding 10 to the set of choices returns a different set of choices:
#|kno>|# (+ {3 4 5} 10)
{13 14 15}
while multiplying the set of choices by itself produces even more options:
#|kno>|# (* {3 4 5} {3 4 5})
{9 12 15 16 15 20 25}
Whenever Kno applies a procedure to a set of choices, it picks each of the choices, applies the procedure, and combines the results; thus, if we define SQUARE as:
(define (square x) (* x x))
and apply it to the same set of choices as above, we get only three choices back, since square is called three times on each single input and that single input is then multipled by itself:
#|kno>|# (square {3 4 5})
{9 16 25}
When a procedure returns a non-deterministic value, we can apply another procedure to it, as in:
#|kno>|# (+ (square {3 4 5}) 10)
{19 26 35}
Most Kno procedures work in exactly this way when given non deterministic sets for arguments, passing on any non-determinism in their arguments to their results. However, some procedures work differently by either returning deterministic results for non-deterministic arguments into a single result or taking deterministic arguments and returning a set of choices (a non- deterministic result).
When a procedure returns a non-deterministic result consisting of one choice, that is the same as a deterministic result. This means that a regular procedure can return a deterministic result from non-deterministic argument, as in:
#|kno>|# (square {3 -3})
9
Built-in procedures for generating deterministic results from non- deterministic inputs include:
-
(pick-one set)
randomly selects one of the choices in set -
(pick-n set *n*)
randomly selects at most n of the choices in set -
(choice->list set)
returns the choices as a list of elements -
(choice-size _set_ )
returns the number of choices in set -
(empty? expr)
or(fail? expr)
returns true if evaluating expr returns no values -
(exists? expr)
returns true if evaluating expr returns any values at all -
(contains? val expr)
returns true the result of evaluating expr includes val
#|kno>|# **(PICK-ONE (CHOICE 2 3 4))**
3
#|kno>|# **(PICK-ONE (CHOICE 2 3 4))**
2
#|kno>|# **(CHOICE-SIZE (CHOICE 2 3 4))**
3
#|kno>|# **(CHOICE-SIZE 8)**
1
#|kno>|# **(CHOICE-SIZE {})**
0
#|kno>|# **(FAIL? (CHOICE))**
#T
#|kno>|# **(FAIL? 3)**
#F
#|kno>|# **(EMPTY? (CHOICE 3 4))**
#F
#|kno>|# **(DEFINE (EVEN? x) (if (zero? (remainder x 2)) x (CHOICE)))**
#|kno>|# **(EXISTS? (CHOICE))**
#F
#|kno>|# **(EXISTS? 3)**
#T
#|kno>|# **(EXISTS? (even? (CHOICE 3 5 9)))**
#F
#|kno>|# **(EXISTS? (even? (CHOICE 2 3 5 9)))**
#T
#|kno>|# **(CONTAINS? 2 (CHOICE 2 3 4))**
#t
#|kno>|# **(CONTAINS? 5 (CHOICE 2 3 4))**
#f
#|kno>|# **(CONTAINS? 8 (+ (CHOICE 2 3 4) (CHOICE 4 5 6)))**
#t
Other built-in procedures generate non-deterministic results from
deterministic arguments. The most basic such procedure is CHOICE
, which
returns its arguments non-deterministically, e.g.
#|kno>|# (choice 3 4 5)
{3 4 5}
#|kno>|# (+ (CHOICE 3 4 5) 10)
{13 14 15}
while another important one is ELTS
which returns the elements of a sequence
non-deterministically, e.g.:
#|kno>|# (elts '(a b c))
{A B C}
#|kno>|# (elts "def")
{#\d #\f #\e}
A procedure can also return no choices at all. This "return value" is called a
failure and is indicated by pair of empty curly braces "{}
", e.g.
#|kno>|# **(CHOICE)**
{}
when a procedure is called on a failure, the procedure itself returns a failure, so:
#|kno>|# **(+ (CHOICE 10 8) (CHOICE))**
{}
This special result, indicating no returned choices, is called a failure because of the way that choices are used in searching by non-deterministic programming. If you think of a given procedure as doing some `search' given the constraints of its arguments, returning the empty choice can be considered as "failing" in the part of the search.
The early termination on failure is called "pruning." We say that the call to
+
was pruned because the second call to CHOICE failed. Note that if a
subexpression fails in this way, none of the remaining arguments are
evaluated, E.G.
#|kno>|# **(+ (CHOICE) (begin (lineout "last argument") 3))**
{}
doesn't produce the output line ``last argument`' because the whole expression is pruned before the final form is evaluated.
Non-deterministic return values can be used to represent sets, as in the following definition of set intersection, which specifies the base case and naturally generalizes:
#|kno>|# **(define (intersect x y) (if (equal? x y) x (fail)))**
#|kno>|# **(intersect (CHOICE 3 4 5 6) (CHOICE 5 6 7 8))**
{5 6}
We can see the value combination process in action by adding trace statements
to the INTERSECT
procedure, as in:
#|kno>|# **(define (intersect x y)
(lineout "INTERSECT " x " = " y " is " (equal? x y))
(if (equal? x y) x {}))**
#|kno>|# **(intersect (CHOICE 3 4 5) (CHOICE 5 6 7))**
INTERSECT 3 = 5 is #f
INTERSECT 3 = 6 is #f
INTERSECT 3 = 7 is #f
INTERSECT 4 = 5 is #f
INTERSECT 4 = 6 is #f
INTERSECT 4 = 7 is #f
INTERSECT 5 = 5 is #t
INTERSECT 5 = 6 is #f
INTERSECT 5 = 7 is #f
5
Of course, this is an inefficient way to compute intersections. Kno provides a number of special forms for dealing with non-deterministic values, which we describe in the next section.
There are a variety of Kno special forms for dealing with non- deterministic values. They are called "special" forms because they do not follow the normal rules for non-deterministic procedure combination.
(INTERSECTION expr1 expr2)
evaluates expr1 and expr2 and returns only the values returned by both expressions. E.G.
#|kno>|# **(INTERSECTION {3 4 5} {2 4 6})**
4
On very large choices, operations like intersection can be very time
(UNION expr1 expr2)
evaluates expr1 and expr2 and returns the results from both. E.G.
#|kno>|# **(UNION {3 4 5} {2 4 6})**
{2 3 4 5 6}
(DIFFERENCE expr1 expr2)
evaluates expr1 and expr2 and returns the results of expr1 which are not returned by expr2. E.G.
#|kno>|# **(DIFFERENCE {3 4 5} {2 4 6})**
{3 5}
(TRY expri...)
Evaluates each expri in order, returning the first one which doesn't fail (e.g. which produces any values at all), E.G.
#|kno>|# **(TRY (INTERSECTION (CHOICE 3 4 5) (CHOICE 6 7 8)) ; _This one fails_
(INTERSECTION (CHOICE 3 4 5) (CHOICE 1 2 3)) ; _This one doesn't_
(INTERSECTION (CHOICE 3 4 5) (CHOICE 4 5 6))) ; _This one doesn't get a chance_**
3
You may have figured out that non-deterministic evaluation and pruning can't apply to the definitions above or else an expression like:
(UNION (CHOICE) (CHOICE 3 4))
would automatically be pruned. Some other special forms also break the default
rules for combination and pruning. For instance, the formatted output
functions such as LINEOUT
don't do automatic enumeration and pruning, so you
get the following behavior:
#|kno>|# **(LINEOUT "This is empty: " (CHOICE) " but this isn't: " (CHOICE 2 3))**
This is empty: {} but this isn't: {2 3}
User procedures (like the procedure INTERSECT
which we defined above)
automatically invoke the interpreter's search and combination mechanisms. For
instance, the following fragment generates possible sentences:
#|kno>|# **(DEFINE (sentence subject verb object) (list subject verb object))**
#|kno>|# **(sentence (CHOICE "Moe" "Larry" "Curly")
(CHOICE "hit" "kissed")
(CHOICE "Huey" "Dewey" "Louie"))**
{("Moe" "hit" "Huey") ("Larry" "hit" "Huey") ("Curly" "hit" "Huey")
("Moe" "kissed" "Huey") ("Larry" "kissed" "Huey")
("Curly" "kissed" "Huey") ("Moe" "hit" "Dewey")
("Larry" "hit" "Dewey") ("Curly" "hit" "Dewey")
("Moe" "kissed" "Dewey") ("Larry" "kissed" "Dewey")
("Curly" "kissed" "Dewey") ("Moe" "hit" "Louie")
("Larry" "hit" "Louie") ("Curly" "hit" "Louie")
("Moe" "kissed" "Louie") ("Larry" "kissed" "Louie")
("Curly" "kissed" "Louie")}
The only caveat to the non-deterministic application of user procedures was
mentioned above. If a user procedure takes a dotted or optional argument, the
argument is bound to a list of the remaining choices rather than a choice
among the lists that they would generate. So, this definition calls LINEOUT
once on the choice {3 4}
:
#|kno>|# **(define (list-choices . x) (lineout "Results are: " (car x)))**
#|kno>|# **(list-choices (CHOICE 3 4))**
Results are: {3 4}
while this definition calls list-choices
separately on the returned values:
#|kno>|# **(define (list-choices x) (lineout "Results are: " x))**
#|kno>|# **(list-choices (CHOICE 3 4))**
Results are: 3
Results are: 4
calls list-choices
separately on the returned values.
The key point is that if a procedure is expecting a choice as an argument and needs the choice to remain a choice (rather than having its elements enumerated), the argument should be extracted from a "dotted" argument. For instance, suppose we wanted to define a function which returned twice the size of a choice, we might try to write it this way:
#|kno>|# **(DEFINE (BAD-DOUBLE-SIZE x) (* 2 (choice-size x)))**
#|kno>|# **(BAD-DOUBLE-SIZE 3)** ; <== this works fine
2
#|kno>|# **(BAD-DOUBLE-SIZE {3 4 5 6})** ; <== this doesn't
2
but that won't work because the x
argument is bound to each of the numbers
in the choice individually, rather than as an entire choice at once. A correct
definition would be:
#|kno>|# **(DEFINE (DOUBLE-SIZE . ARGS) (* 2 (SET-SIZE (CAR ARGS))))**
#|kno>|# **(DOUBLE-SIZE 3)** ; <== this still works fine
2
#|kno>|# **(DOUBLE-SIZE {3 4 5 6})** ; <== and so does this...
8
Choices can be stored and saved in a variety of ways. For instance, the
special form SET!
sets a variable to contain a set of possible values, so
one can say:
#|kno>|# **(SET! small-primes (CHOICE 2 3 5 7 11 13 17 19))**
#|kno>|# **(define (divides? x y) (if (zero? (remainder x y)) y {}))**
#|kno>|# **(divides? 15 small-primes)**
{3 5}
The SET+!
adds a set of values non-deterministically to a variable. For
example,
#|kno>|# **(SET! small-odd-numbers (CHOICE 1 3 5 7))**
#|kno>|# **small-odd-numbers**
{5 1 7 3}
#|kno>|# **(SET+! small-odd-numbers (CHOICE 9 17))**
#|kno>|# **small-odd-numbers**
{5 9 1 7 17 3}
The binding special forms LET
and LET*
can be used to store non-
deterministic values in the same way as set!
. E.G.
#|kno>|# **(define (divides? x y) (if (zero? (remainder x y)) y {}))**
#|kno>|# **(let ((small-primes (CHOICE 2 3 5 7 11 13 17 19)))
(divides? 15 small-primes))**
{3 5}
(For old-time Scheme aficianados, this interpretation of LET
breaks the
equivalence of LET
and LAMBDA
, since writing the above as an application
of a lambda would automatically iterate through the choices..)
Sometimes it is important to be able to process each element of a choice
separately. Kno provides three special forms supporting this kind of
processing, DO-CHOICES, FOR-CHOICES,
and FILTER-CHOICES
:
(DO-CHOICES (var val-expr) expr1 expr2...)
Evaluates all of the expri with var bound to each of the values returned by val-expr. E.G.
#|kno>|# **(DO-CHOICES (x (CHOICE 3 4)) (lineout "I saw a " x))**
I saw a 4
I saw a 3
DO-CHOICES
can be used to unpack a set of values to pass to forms which
don't automatically unpack their arguments (such as LINEOUT
), as in the
following definition which puts each of the values returned by FGET
on a
different line:
(define (print-slot-values frame slot)
(let ((values (fget frame slot)))
(lineout "The " slot " of " frame " is:")
(do-choices (value values)
(lineout " " value))))
(FOR-CHOICES (var val-expr) expr1 expr2...)
Like DO-CHOICES, but combines the results of evaluating the last expri for each value, E.G.
#|kno>|# **(FOR-CHOICES (x (CHOICE 3 4 5 6)) (if (zero? (remainder x 2)) (+ x 3)))**
{7 9}
(FILTER-CHOICES (var value-expr) test-expri...)
Evaluates value-expr and and binds var to each element and returning those elements for which every test-expri returns true given the binding. E.G.
#|kno>|# **(DEFINE (EVEN? x) (zero? (remainder x 2)))**
#|kno>|# **(FILTER-CHOICES (num (CHOICE 1 2 3 4 5 6))
(EVEN? x))**
{2 4 6}
#|kno>|# **(FILTER-CHOICES (num (CHOICE 1 2 3 4 5 6))
(EVEN? num)
( < num 6))**
{2 4}