在《Jess in Action——Java中基于规则的系统》(10多年前写的;我认为Drools是今天要使用的系统?),Ernest Friedman-Hill使用Jess解决了下面给出的约束问题,Jess是一个用Java编写的OPS5风格的前向链生产系统。我想用Prolog解决它。
问题是:我是否正确解决它?
问题所在
四人组高尔夫球手站在发球台上,从左到一排 右。每个高尔夫球手都穿不同颜色的裤子;一个穿着红色 裤子。美联储右边的高尔夫球手穿着蓝色裤子。乔 排在第二位。鲍勃穿着格子裤。汤姆不在位置上 一四个,他没有穿那条丑陋的橙色裤子。
在什么 顺序是四位高尔夫球手开球,每个高尔夫球手的颜色是什么 裤子?
这是斑马拼图的一个实例。另请参阅此演示文稿,了解更复杂的解决方案的精美插图解决方案。
使用杰斯,欧内斯特·弗里德曼-希尔
使用 Jess 生产系统,代码如下。这是来自上述书籍,为清楚起见,变量已重命名。
工作记忆中充满了从高尔夫球手到他们可能的位置和裤子颜色的 32 个链接。find-solution
规则为满足约束的链接集触发。
这似乎很难思考,因为人们不会测试"可能的世界"是否满足约束,而是选择一组满足约束的链接。不清楚这是否确实是一个人正在寻找的。
;; Templates for working memory, basically the links golfer<->pantscolor,
;; and golfer<->position.
(deftemplate pants-color (slot of) (slot is))
(deftemplate position (slot of) (slot is))
;; Generate all possible 'pants-color' and 'position' facts
;; 4 names, each with 4 pants-color: 16 entries
;; 4 names, each with 4 positions: 16 entries
;; This gives the 32 facts describing the links
(defrule generate-possibilities
=>
(foreach ?name (create$ Fred Joe Bob Tom)
(foreach ?color (create$ red blue plaid orange)
(assert (pants-color (of ?name) (is ?color))))
(foreach ?position (create$ 1 2 3 4)
(assert (position (of ?name) (is ?position))))))
;; The “find solution” rule forward-chains and prints out a solution
(defrule find-solution
;; There is a golfer named Fred, whose position is ?p_fred and
;; pants color is ?c_fred
(position (of Fred) (is ?p_fred))
(pants-color (of Fred) (is ?c_fred))
;; The golfer to Fred's immediate right (who is not Fred) is wearing
;; blue pants.
(position (of ?n&~Fred) (is ?p&:(eq ?p (+ ?p_fred 1))))
(pants-color (of ?n&~Fred) (is blue&~?c_fred))
;; Joe is in position #2
(position (of Joe) (is ?p_joe&2&~?p_fred))
(pants-color (of Joe) (is ?c_joe&~?c_fred))
;; Bob is wearing the plaid pants (so his position is not “n” either
;; because “n” has blue pants)
(position (of Bob) (is ?p_bob&~?p_fred&~?n&~?p_joe))
(pants-color (of Bob&~?n) (is plaid&?c_bob&~?c_fred&~?c_joe))
;; Tom isn't in position 1 or 4 and isn't wearing orange (and not blue
;; either)
(position (of Tom&~?n) (is ?p_tom&~1&~4&~?p_fred&~?p_joe&~?p_bob))
(pants-color (of Tom) (is ?c_tom&~orange&~blue&~?c_fred&~?c_joe&~?c_bob))
=>
(printout t Fred " " ?p_fred " " ?c_fred crlf)
(printout t Joe " " ?p_joe " " ?c_joe crlf)
(printout t Bob " " ?p_bob " " ?c_bob crlf)
(printout t Tom " " ?p_tom " " ?c_tom crlf crlf))
我在Prolog中的第一个解决方案
事实证明,这是不优雅和严厉的(见其他答案)
让我们寻找一个数据结构来描述解决方案,如下所示:选择一个列表,在每个位置都有一个具有"名称"和"裤子颜色"的"高尔夫球手":[golfer(N0,C0),golfer(N1,C1),golfer(N2,C2),golfer(N3,C3)]
.每个高尔夫球手的发球位置也由列表中的实际位置给出从 0 到 3;该职位没有像golfer(Name,Color,Position)
那样明确给出。
solution(L) :-
% select possible pants colors which must be pairwise different; for
% fast fail, we check often
is_pants_color(C0),
is_pants_color(C1),are_pairwise_different([C0,C1]),
is_pants_color(C2),are_pairwise_different([C0,C1,C2]),
is_pants_color(C3),are_pairwise_different([C0,C1,C2,C3]),
% select possible golfer names which must be pairwise different; for
% fast fail, we check often
is_name(N0),
% we know that joe is second in line, so we can plonck that condition
% in here immediately
N1 = joe,
is_name(N1),are_pairwise_different([N0,N1]),
is_name(N2),are_pairwise_different([N0,N1,N2]),
is_name(N3),are_pairwise_different([N0,N1,N2,N3]),
% instantiate the solution in a unique order (we don't change the order
% as we permute exhuastively permute colors and names)
L = [golfer(N0,C0),golfer(N1,C1),golfer(N2,C2),golfer(N3,C3)],
% tom is not in position one or four; express this clearly using
% "searchWithPosition" instead of implicitly by unification with L
search(tom,L,golfer(_,_,TomPosition)),
TomPosition == 0,
TomPosition == 3,
% check additional constraints using L
rightOf(fred,L,golfer(_,blue)),
search(bob,L,golfer(_,plaid,_)),
+search(tom,L,golfer(_,hideous_orange,_)).
% here we stipulate the colors
is_pants_color(red).
is_pants_color(blue).
is_pants_color(plaid).
is_pants_color(hideous_orange).
% here we stipulate the names
is_name(joe).
is_name(bob).
is_name(tom).
is_name(fred).
% helper predicate
are_pairwise_different(L) :- sort(L,LS), length(L,Len), length(LS,Len).
% Search a golfer by name in the solution list, iteratively.
% Also return the position 0..3 for fun and profit (allows to express the
% constraint on the position)
% We "know" that names are unique, so cut on the first clause.
search(Name,L,golfer(Name,C,Pos)) :-
searchWithPosition(Name,L,golfer(Name,C,Pos),0).
searchWithPosition(Name,[golfer(Name,C)|_],golfer(Name,C,Pos),Pos) :- !.
searchWithPosition(Name,[_|R],golfer(Name,C,PosOut),PosIn) :-
PosDown is PosIn+1, searchWithPosition(Name,R,golfer(Name,C,PosOut),PosDown).
% Search the golfer to the right of another golfer by name in the list,
% iteratively. We "know" that names are unique, so cut on the first clause
rightOf(Name,[golfer(Name,_),golfer(N,C)|_],golfer(N,C)) :- !.
rightOf(Name,[_|R],golfer(N,C)) :- rightOf(Name,R,golfer(N,C)).
让我们运行这个:
?:- solution(L).
L = [golfer(fred, hideous_orange),
golfer(joe, blue),
golfer(tom, red),
golfer(bob, plaid)]
紧凑型解决方案
golfers(S) :-
length(G, 4),
choices([
g(1, _, _),
g(2, joe, _), % Joe is second in line.
g(3, _, _),
g(4, _, _),
g(_, _, orange),
g(_, _, red), % one is wearing red pants
g(_, bob, plaid), % Bob is wearing plaid pants
g(P, fred, _), % The golfer to Fred’s immediate right
g(Q, _, blue), % ....is wearing blue pants
g(Pos, tom, Pants) % Tom isn’t in position one or four, and
% ... he isn’t wearing the orange pants
], G),
Q is P+1,
Pos = 1, Pos = 4, Pants = orange, sort(G,S).
choices([],_).
choices([C|Cs],G) :- member(C,G), choices(Cs,G).
OP添加的说明:为什么有效
- 使用
length/2
创建包含 4 个未初始化元素的列表 G - 对于传递给
choices/2
的第一个参数中的每个元素 C,确保 C 是 G 的成员。- 前 4 个条目将按顺序分配(希望是确定性的),并且由于它们无法统一,这将导致在第 4 次调用
member/2
之后出现类似[g(1, _G722, _G723), g(2, joe, _G730), g(3, _G736, _G737), g(4, _G743, _G744)]
的内容。 - 在
choices/2
返回之后,G 被统一为一个结构,该结构满足传递给choices/2
的约束列表中的每个约束,特别是:- 列出位置 1,2,3,4
- 列出乔、鲍勃、弗雷德、汤姆的名字
- 颜色橙色,格子,红色,蓝色列出
- 。这意味着我们甚至不必检查颜色、名称或位置是否出现两次 - 它只能只出现一次。
- 其他约束无法传递给
choices/2
(没有办法说g(P, fred, _), g(P+1, _, blue), g(not-in{1,4}, tom, not-in{orange})
之类的话并将其传递给choices/2
)。因此,这些附加约束是通过与 G 内容统一的变量来检查的。 - 如果这些附加约束失败,将发生
choices/2
回溯,从而超过member/2
。此时堆栈上有 9 个member/2
调用,这将被详尽地尝试,尽管回溯过去的成员分配g(4, _, _)
没有用。 - 一旦找到可接受的解决方案,就会对其进行排序,程序就会成功。
- 前 4 个条目将按顺序分配(希望是确定性的),并且由于它们无法统一,这将导致在第 4 次调用
紧凑型解决方案,经过修改
由OP添加:
以上表明,略有改善是可能的。该程序在第一个解决方案之后找不到任何其他(相同的)解决方案:
golfers(G) :-
G=[g(1,_,_),g(2,_,_),g(3,_,_),g(4,_,_)],
choices([
g(2, joe, _), % Joe is second in line.
g(_, _, orange),
g(_, _, red), % one is wearing red pants
g(_, bob, plaid), % Bob is wearing plaid pants
g(P, fred, _), % The golfer to Fred’s immediate right is
g(Q, _, blue), % ...wearing blue pants
g(Pos, tom, Pants) % Tom isn’t in position one or four, and
% ...he isn’t wearing the hideous orange pants
], G),
Q is P+1,
Pos = 1, Pos = 4, Pants = orange.
choices([],_).
choices([C|Cs],G) :- member(C,G), choices(Cs,G).
为什么有效
- 立即定义生成的 G 的结构,而不是使用"长度"创建四个未知元素的列表
- 在这个"proto-G"中,列表元素是按位置自然排序的;我们不会找到不同的解决方案,其中
g(P,_,_)
按位置排列。- 因此,我们可以摆脱
g(1,_,_), g(3,_,_), g(4,_,_)
约束 - 如果还想确保名称和颜色只使用一次(这不是必需的,因为构造必须如此),则可以使用
g(1,N1,C1), g(2,N2,C2), g(3,N3,C3), g(4,N4,C4)
通过choices/2
捕获名称和颜色,并通过sort/2
确保Ni和Ci是唯一的:sort([N1,N2,N3,N4],[bob,fred,joe,tom]), sort([C1,C2,C3,C4],[blue,orange,plaid,red])
- 因此,我们可以摆脱
另一种解决方案
Prolog使编写"语言"变得容易。让我们声明问题,并制作一个微型DSL来解决:
golfers_pants([G1,G2,G3,G4]) :-
maplist(choice([G1,G2,G3,G4]),[
% my note: we are going to compute on positions, so fill the 'column' with domain values
g(1, _, _),
% Joe is second in line.
g(2, joe, _),
g(3, _, _),
g(4, _, _),
% my note: someone is wearing 'hideous orange pants' not mentioned positively elsewhere
g(_, _, orange),
% one is wearing red pants
g(_, _, red),
% Bob is wearing plaid pants
g(_, bob, plaid),
% The golfer to Fred’s immediate right is wearing blue pants
g(P, fred, _), g(Q, _, blue), Q is P+1,
% Tom isn’t in position one or four, and he isn’t wearing the hideous orange pants
g(Pos, tom, Pants), Pos = 1, Pos = 4, Pants = orange
]).
choice(G,C) :- C = g(_,_,_) -> member(C,G) ; call(C).
Jess解决方案,在Prolog中重写
这是为了完成。
在SWI Prolog中重写Jess解决方案(但不是在SWISH中,因为我们现在使用assert
)表明:
- 有很多详尽的枚举在"引擎盖下"进行
- 正向链接生产系统可能不是解决这种"有限搜索空间上的约束满足"问题的最佳工具
- 规则条件可能会从一些概念清理中受益
所以,让我们直接翻译一下:
% Define the possible names, colors and positions
names([fred,joe,bob,tom]).
colors([red,blue,plaid,orange]).
positions([1,2,3,4]).
run :- names(Ns),
colors(Cs),
positions(Ps),
fill_working_memory(pantscolor,Ns,Cs),
fill_working_memory(position,Ns,Ps).
fireable(SortedResult) :-
position(fred,P_fred),
pantscolor(fred,C_fred),
position(N,P) , N == fred,
P is P_fred+1,
pantscolor(N,blue) , N == fred,
+member(C_fred,[blue]),
position(joe,P_joe) , P_joe == 2,
+member(P_joe,[P_fred]),
pantscolor(joe,C_joe) , +member(C_joe,[C_fred]),
position(bob, P_bob) , +member(P_bob,[P_fred,N,P_joe]),
pantscolor(bob, C_bob), N == bob,
C_bob = plaid,
+member(C_bob, [C_fred,C_joe]),
position(tom, P_tom) , N == tom,
+member(P_tom,[1,4,P_fred,P_joe,P_bob]),
pantscolor(tom, C_tom), +member(C_tom,[orange,blue,C_fred,C_joe,C_bob]),
% build clean result
Result = [g(P_fred,fred,C_fred),
g(P_bob,bob,C_bob),
g(P_joe,joe,C_joe),
g(P_tom,tom,C_tom)],
sort(Result,SortedResult).
% -- Helper to assert initial facts into the working memory
fill_working_memory(PredSym,Ns,Vs) :-
product(Ns,Vs,Cartesian),
forall(member([N,V], Cartesian), factify(PredSym,N,V)).
factify(PredSym,N,V) :- Term=..([PredSym,N,V]), writeln(Term), assertz(Term).
% -- These should be in a library somewhere --
% Via https://gist.github.com/raskasa/4282471
% pairs(+N,+Bs,-Cs)
% returns in Cs the list of pairs [N,any_element_of_B]
pairs(_,[],[]) :- !.
pairs(N,[B|Bs],[[N,B]|Cs]) :- pairs(N,Bs,Cs).
% product(+As,+Bs,-Cs)
% returns in Cs the cartesian product of lists As and Bs
% product([x,y], [a,b,c], [[x, a], [x, b], [x, c], [y, a], [y, b], [y, c]])
% Would be interesting to make this a product(+As,+Bs,?Cs)
product([],_,[]) :- !.
product([A|As],Bs,Cs) :- pairs(A,Bs,Xs),
product(As,Bs,Ys),
append(Xs,Ys,Cs).
让我们运行这个:
?- run, fireable(X).
X = [g(1, fred, orange),
g(2, joe, blue),
g(3, tom, red),
g(4, bob, plaid)] .
出于某种原因,swipl
在第 5 次执行左右后变得狗慢。垃圾收集开始了?