解决XKCD中的NP完全问题
有问题的问题/漫画:http://xkcd.com/287/
我不确定这是否是最好的方法,但是到目前为止,这是我要提出的。我正在使用CFML,但任何人都应该可以读取。
<cffunction name="testCombo" returntype="boolean"> <cfargument name="currentCombo" type="string" required="true" /> <cfargument name="currentTotal" type="numeric" required="true" /> <cfargument name="apps" type="array" required="true" /> <cfset var a = 0 /> <cfset var found = false /> <cfloop from="1" to="#arrayLen(arguments.apps)#" index="a"> <cfset arguments.currentCombo = listAppend(arguments.currentCombo, arguments.apps[a].name) /> <cfset arguments.currentTotal = arguments.currentTotal + arguments.apps[a].cost /> <cfif arguments.currentTotal eq 15.05> <!--- print current combo ---> <cfoutput><strong>#arguments.currentCombo# = 15.05</strong></cfoutput><br /> <cfreturn true /> <cfelseif arguments.currentTotal gt 15.05> <cfoutput>#arguments.currentCombo# > 15.05 (aborting)</cfoutput><br /> <cfreturn false /> <cfelse> <!--- less than 15.05 ---> <cfoutput>#arguments.currentCombo# < 15.05 (traversing)</cfoutput><br /> <cfset found = testCombo(arguments.currentCombo, arguments.currentTotal, arguments.apps) /> </cfif> </cfloop> </cffunction> <cfset mf = {name="Mixed Fruit", cost=2.15} /> <cfset ff = {name="French Fries", cost=2.75} /> <cfset ss = {name="side salad", cost=3.35} /> <cfset hw = {name="hot wings", cost=3.55} /> <cfset ms = {name="moz sticks", cost=4.20} /> <cfset sp = {name="sampler plate", cost=5.80} /> <cfset apps = [ mf, ff, ss, hw, ms, sp ] /> <cfloop from="1" to="6" index="b"> <cfoutput>#testCombo(apps[b].name, apps[b].cost, apps)#</cfoutput> </cfloop>
上面的代码告诉我,唯一加起来为$ 15.05的组合是7个混合水果订单,并且需要232次testCombo函数的执行才能完成。
是否有更好的算法可以得出正确的解决方案?我找到正确的解决方案了吗?
解决方案
阅读背包问题。
实际上,我已经对算法进行了一些重构。我错过了几个正确的组合,这是由于这样的事实,当费用超过15.05时我就立即返回-我不费心检查我可以添加的其他(便宜的)物品。这是我的新算法:
<cffunction name="testCombo" returntype="numeric"> <cfargument name="currentCombo" type="string" required="true" /> <cfargument name="currentTotal" type="numeric" required="true" /> <cfargument name="apps" type="array" required="true" /> <cfset var a = 0 /> <cfset var found = false /> <cfset var CC = "" /> <cfset var CT = 0 /> <cfset tries = tries + 1 /> <cfloop from="1" to="#arrayLen(arguments.apps)#" index="a"> <cfset combos = combos + 1 /> <cfset CC = listAppend(arguments.currentCombo, arguments.apps[a].name) /> <cfset CT = arguments.currentTotal + arguments.apps[a].cost /> <cfif CT eq 15.05> <!--- print current combo ---> <cfoutput><strong>#CC# = 15.05</strong></cfoutput><br /> <cfreturn true /> <cfelseif CT gt 15.05> <!--<cfoutput>#arguments.currentCombo# > 15.05 (aborting)</cfoutput><br />--> <cfelse> <!--- less than 15.50 ---> <!--<cfoutput>#arguments.currentCombo# < 15.05 (traversing)</cfoutput><br />--> <cfset found = testCombo(CC, CT, arguments.apps) /> </cfif> </cfloop> <cfreturn found /> </cffunction> <cfset mf = {name="Mixed Fruit", cost=2.15} /> <cfset ff = {name="French Fries", cost=2.75} /> <cfset ss = {name="side salad", cost=3.35} /> <cfset hw = {name="hot wings", cost=3.55} /> <cfset ms = {name="moz sticks", cost=4.20} /> <cfset sp = {name="sampler plate", cost=5.80} /> <cfset apps = [ mf, ff, ss, hw, ms, sp ] /> <cfset tries = 0 /> <cfset combos = 0 /> <cfoutput> <cfloop from="1" to="6" index="b"> #testCombo(apps[b].name, apps[b].cost, apps)# </cfloop> <br /> tries: #tries#<br /> combos: #combos# </cfoutput>
输出:
Mixed Fruit,Mixed Fruit,Mixed Fruit,Mixed Fruit,Mixed Fruit,Mixed Fruit,Mixed Fruit = 15.05 Mixed Fruit,hot wings,hot wings,sampler plate = 15.05 Mixed Fruit,hot wings,sampler plate,hot wings = 15.05 Mixed Fruit,sampler plate,hot wings,hot wings = 15.05 false false false hot wings,Mixed Fruit,hot wings,sampler plate = 15.05 hot wings,Mixed Fruit,sampler plate,hot wings = 15.05 hot wings,hot wings,Mixed Fruit,sampler plate = 15.05 hot wings,sampler plate,Mixed Fruit,hot wings = 15.05 false false sampler plate,Mixed Fruit,hot wings,hot wings = 15.05 sampler plate,hot wings,Mixed Fruit,hot wings = 15.05 false tries: 2014 combos: 12067
我认为这可能具有所有正确的组合,但我的问题仍然存在:是否有更好的算法?
现在我们已经拥有了所有正确的组合,但是我们仍然需要检查的内容比需要的要多(结果显示的许多排列证明了这一点)。另外,我们将省略最后一个达到15.05标记的项目。
我有一个PHP版本,可以进行209次递归调用的迭代(如果得到所有排列,则可以在2012年实现)。如果恰好在循环结束之前拔出刚刚检查的项目,则可以减少计数。
我不知道CF语法,但是会是这样的:
<cfelse> <!--- less than 15.50 ---> <!--<cfoutput>#arguments.currentCombo# < 15.05 (traversing)</cfoutput><br />--> <cfset found = testCombo(CC, CT, arguments.apps) /> ------- remove the item from the apps array that was just checked here ------ </cfif> </cfloop>
编辑:供参考,这是我的PHP版本:
<? function rc($total, $string, $m) { global $c; $m2 = $m; $c++; foreach($m as $i=>$p) { if ($total-$p == 0) { print "$string $i\n"; return; } if ($total-$p > 0) { rc($total-$p, $string . " " . $i, $m2); } unset($m2[$i]); } } $c = 0; $m = array("mf"=>215, "ff"=>275, "ss"=>335, "hw"=>355, "ms"=>420, "sp"=>580); rc(1505, "", $m); print $c; ?>
输出
mf mf mf mf mf mf mf mf hw hw sp 209
编辑2:
由于要解释为什么可以删除元素将花费比我在注释中多得多的内容,因此在此处添加它。
基本上,每次递归都会找到包含当前搜索元素的所有组合(例如,第一步将找到包含至少一个混合水果的所有内容)。理解它的最简单方法是跟踪执行,但是由于这将占用大量空间,因此我将其视为目标为6.45.
MF (2.15) MF (4.30) MF (6.45) * FF (7.05) X SS (7.65) X ... [MF removed for depth 2] FF (4.90) [checking MF now would be redundant since we checked MF/MF/FF previously] FF (7.65) X ... [FF removed for depth 2] SS (5.50) ... [MF removed for depth 1]
至此,我们已经检查了包含任何混合水果的每个组合,因此无需再次检查混合水果。我们还可以使用相同的逻辑在每个更深层次的递归中修剪数组。
像这样跟踪它实际上暗示了另一个节省时间的方法-知道价格是从低到高排序的,这意味着一旦超过目标,我们就无需继续检查商品。
这是F#的解决方案:
#light type Appetizer = { name : string; cost : int } let menu = [ {name="fruit"; cost=215} {name="fries"; cost=275} {name="salad"; cost=335} {name="wings"; cost=355} {name="moz sticks"; cost=420} {name="sampler"; cost=580} ] // Choose: list<Appetizer> -> list<Appetizer> -> int -> list<list<Appetizer>> let rec Choose allowedMenu pickedSoFar remainingMoney = if remainingMoney = 0 then // solved it, return this solution [ pickedSoFar ] else // there's more to spend [match allowedMenu with | [] -> yield! [] // no more items to choose, no solutions this branch | item :: rest -> if item.cost <= remainingMoney then // if first allowed is within budget, pick it and recurse yield! Choose allowedMenu (item :: pickedSoFar) (remainingMoney - item.cost) // regardless, also skip ever picking more of that first item and recurse yield! Choose rest pickedSoFar remainingMoney] let solutions = Choose menu [] 1505 printfn "%d solutions:" solutions.Length solutions |> List.iter (fun solution -> solution |> List.iter (fun item -> printf "%s, " item.name) printfn "" ) (* 2 solutions: fruit, fruit, fruit, fruit, fruit, fruit, fruit, sampler, wings, wings, fruit, *)
递归(在Perl中)是否更优雅?
#!/usr/bin/perl use strict; use warnings; my @weights = (2.15, 2.75, 3.35, 3.55, 4.20, 5.80); my $total = 0; my @order = (); iterate($total, @order); sub iterate { my ($total, @order) = @_; foreach my $w (@weights) { if ($total+$w == 15.05) { print join (', ', (@order, $w)), "\n"; } if ($total+$w < 15.05) { iterate($total+$w, (@order, $w)); } } }
输出
`marco @ unimatrix-01:〜$ ./xkcd-knapsack.pl
2.15、2.15、2.15、2.15、2.15、2.15、2.15
2.15、3.55、3.55、5.8
2.15、3.55、5.8、3.55
2.15、5.8、3.55、3.55
3.55、2.15、3.55、5.8
3.55、2.15、5.8、3.55
3.55、3.55、2.15、5.8
3.55、5.8、2.15、3.55
5.8、2.15、3.55、3.55
5.8、3.55、2.15、3.55
`
关于NP完全问题的意义不在于在一个小的数据集上是棘手的,而是解决它的工作量以大于多项式的速率增长,即没有O(n ^ x)算法。
如果时间复杂度为O(n!),如(我相信)上面提到的两个问题,那就是NP。
从@rcar的答案中学习,并在以后进行另一次重构,我得到了以下内容。
与我编写的许多代码一样,我已经从CFML重构为CFScript,但是代码基本上是相同的。
我添加了他关于数组动态起点的建议(而不是通过值传递数组并更改其值以用于将来的递归),这使我获得了与他得到的相同的统计信息(209个递归,571个组合价格检查(循环迭代) )),然后对此进行了改进,即假设该数组将按成本进行排序-因为它是-且一旦超过目标价格便立即中断。有了休息,我们减少到209个递归和376个循环迭代。
function testCombo(minIndex, currentCombo, currentTotal){ var a = 0; var CC = ""; var CT = 0; var found = false; tries += 1; for (a=arguments.minIndex; a <= arrayLen(apps); a++){ combos += 1; CC = listAppend(arguments.currentCombo, apps[a].name); CT = arguments.currentTotal + apps[a].cost; if (CT eq 15.05){ //print current combo WriteOutput("<strong>#CC# = 15.05</strong><br />"); return(true); }else if (CT gt 15.05){ //since we know the array is sorted by cost (asc), //and we've already gone over the price limit, //we can ignore anything else in the array break; }else{ //less than 15.50, try adding something else found = testCombo(a, CC, CT); } } return(found); } mf = {name="mixed fruit", cost=2.15}; ff = {name="french fries", cost=2.75}; ss = {name="side salad", cost=3.35}; hw = {name="hot wings", cost=3.55}; ms = {name="mozarella sticks", cost=4.20}; sp = {name="sampler plate", cost=5.80}; apps = [ mf, ff, ss, hw, ms, sp ]; tries = 0; combos = 0; testCombo(1, "", 0); WriteOutput("<br />tries: #tries#<br />combos: #combos#");
可以对算法进行哪些其他改进?
即使背包是NP完整版,它也是一个非常特殊的问题:常用的动态程序实际上非常出色(http://en.wikipedia.org/wiki/Knapsack_problem)
并且,如果我们进行了正确的分析,结果将是O(nW),n是项的of,W是目标数。问题是当我们必须在较大的W上进行DP运算时,这就是我们得到NP行为的时候。但是,在大多数情况下,背包的行为都相当合理,我们可以毫无问题地解决背包中的很多问题。就NP完全问题而言,背包是最简单的问题之一。
item(X) :- member(X,[215, 275, 335, 355, 420, 580]). solution([X|Y], Z) :- item(X), plus(S, X, Z), Z >= 0, solution(Y, S). solution([], 0).
它提供了解决方案的所有排列方式,但是我认为我在代码大小方面击败了其他所有人。
?- solution(X, 1505). X = [215, 215, 215, 215, 215, 215, 215] ; X = [215, 355, 355, 580] ; X = [215, 355, 580, 355] ; X = [215, 580, 355, 355] ; X = [355, 215, 355, 580] ; X = [355, 215, 580, 355] ; X = [355, 355, 215, 580] ; X = [355, 355, 580, 215] ; X = [355, 580, 215, 355] ; X = [355, 580, 355, 215] ; X = [580, 215, 355, 355] ; X = [580, 355, 215, 355] ; X = [580, 355, 355, 215] ; No
swiprolog解决方案:
>>> from constraint import * >>> problem = Problem() >>> menu = {'mixed-fruit': 2.15, ... 'french-fries': 2.75, ... 'side-salad': 3.35, ... 'hot-wings': 3.55, ... 'mozarella-sticks': 4.20, ... 'sampler-plate': 5.80} >>> for appetizer in menu: ... problem.addVariable( appetizer, [ menu[appetizer] * i for i in range( 8 )] ) >>> problem.addConstraint(ExactSumConstraint(15.05)) >>> problem.getSolutions() [{'side-salad': 0.0, 'french-fries': 0.0, 'sampler-plate': 5.7999999999999998, 'mixed-fruit': 2.1499999999999999, 'mozarella-sticks': 0.0, 'hot-wings': 7.0999999999999996}, {'side-salad': 0.0, 'french-fries': 0.0, 'sampler-plate': 0.0, 'mixed-fruit': 15.049999999999999, 'mozarella-sticks': 0.0, 'hot-wings': 0.0}]
这是使用constraint.py的解决方案
因此,解决方案是订购一个采样盘,一个混合水果和2个热翅,或者订购7个混合水果。
.... private void findAndReportSolutions( int target, // goal to be achieved int balance, // amount of goal remaining int index // menu item to try next ) { ++calls; if (balance == 0) { reportSolution(target); return; // no addition to perfect order is possible } if (index == items.length) { ++falls; return; // ran out of menu items without finding solution } final int price = items[index].price; if (balance < price) { return; // all remaining items cost too much } int maxCount = balance / price; // max uses for this item for (int n = maxCount; 0 <= n; --n) { // loop for this item, recur for others ++loops; counts[index] = n; findAndReportSolutions( target, balance - n * price, index + 1 ); } } public void reportSolutionsFor(int target) { counts = new int[items.length]; calls = loops = falls = 0; findAndReportSolutions(target, target, 0); ps.printf("%d calls, %d loops, %d falls%n", calls, loops, falls); } public static void main(String[] args) { MenuItem[] items = { new MenuItem("mixed fruit", 215), new MenuItem("french fries", 275), new MenuItem("side salad", 335), new MenuItem("hot wings", 355), new MenuItem("mozzarella sticks", 420), new MenuItem("sampler plate", 580), }; Solver solver = new Solver(items); solver.reportSolutionsFor(1505); } ...
我会对算法本身的设计提出一个建议(这就是我如何解释原始问题的意图)。这是我编写的解决方案的一部分:
(请注意,构造函数会按价格递增对菜单项进行排序,以在剩余余额小于任何剩余菜单项时启用固定时间提前终止。)
7 mixed fruit (1505) = 1505 1 mixed fruit (215) + 2 hot wings (710) + 1 sampler plate (580) = 1505 348 calls, 347 loops, 79 falls
样本运行的输出为:
我要强调的设计建议是,在上面的代码中,每个findAndReportSolution(...)嵌套(递归)调用都处理一个菜单项的数量,该数量由" index"参数标识。换句话说,递归嵌套与一组嵌套循环的行为并行。最外层计算第一个菜单项的可能用途,下一个计数第二个菜单项的用途,依此类推。(当然,递归的使用使代码摆脱了对特定数量菜单项的依赖!)
我建议这样做可以使代码设计更容易,并且更容易理解每次调用的功能(考虑到特定项目的所有可能使用,将菜单的其余部分委派给下级调用)。这也避免了产生多项目解决方案的所有安排的组合爆炸(如上述输出的第二行,仅发生一次,而不是以项目的不同顺序重复发生)。
我试图最大化代码的"显而易见性",而不是试图最小化某些特定方法的调用次数。例如,上面的设计让委派的调用确定是否已达到解决方案,而不是将检查结果包装在调用点周围,这样可以减少调用次数,但会使代码混乱。
嗯,你知道这很奇怪。解决方案是菜单上第一项中的七个。
既然这显然是要在短时间内用纸和笔解决的,为什么不将订单总数除以每件商品的价格,以查看是否有机会订购了一件商品的倍数?
例如,
15.05 / 2.15 = 7种混合水果
15.05 / 2.75 = 5.5炸薯条。
然后继续简单的组合...
15 /(2.15 + 2.75)= 3.06122449混合水果和炸薯条。
换句话说,假设该解决方案被认为是简单且可解决的,而无需访问计算机。然后测试最简单,最明显(因此隐藏在视线中)的解决方案是否有效。
我发誓我会在这个周末在当地的购物街上拉这个,当我在俱乐部关门后的4:30 AM订购价值$ 4.77的开胃菜(含税)。
这是Clojure中的并发实现。要计算"(价格为15.05的商品)",大约需要进行14次组合生成的递归,并进行大约10次可能性检查。花了我大约6分钟的时间,才能在我的Intel Q9300上计算出"(含价格的商品100)"。
;; np-complete.clj ;; A Clojure solution to XKCD #287 "NP-Complete" ;; By Sam Fredrickson ;; ;; The function "items-with-price" returns a sequence of items whose sum price ;; is equal to the given price, or nil. (defstruct item :name :price) (def *items* #{(struct item "Mixed Fruit" 2.15) (struct item "French Fries" 2.75) (struct item "Side Salad" 3.35) (struct item "Hot Wings" 3.55) (struct item "Mozzarella Sticks" 4.20) (struct item "Sampler Plate" 5.80)}) (defn items-with-price [price] (let [check-count (atom 0) recur-count (atom 0) result (atom nil) checker (agent nil) ; gets the total price of a seq of items. items-price (fn [items] (apply + (map #(:price %) items))) ; checks if the price of the seq of items matches the sought price. ; if so, it changes the result atom. if the result atom is already ; non-nil, nothing is done. check-items (fn [unused items] (swap! check-count inc) (if (and (nil? @result) (= (items-price items) price)) (reset! result items))) ; lazily generates a list of combinations of the given seq s. ; haven't tested well... combinations (fn combinations [cur s] (swap! recur-count inc) (if (or (empty? s) (> (items-price cur) price)) '() (cons cur (lazy-cat (combinations (cons (first s) cur) s) (combinations (cons (first s) cur) (rest s)) (combinations cur (rest s))))))] ; loops through the combinations of items, checking each one in a thread ; pool until there are no more combinations or the result atom is non-nil. (loop [items-comb (combinations '() (seq *items*))] (if (and (nil? @result) (not-empty items-comb)) (do (send checker check-items (first items-comb)) (recur (rest items-comb))))) (await checker) (println "No. of recursions:" @recur-count) (println "No. of checks:" @check-count) @result))
这只会给出第一个找到的答案,如果没有答案,则为nil,因为这就是所有问题的要求。为什么要求我们做更多的工作;)?
class Solver(object): def __init__(self): self.solved = False self.total = 0 def solve(s, p, pl, curList = []): poss = [i for i in sorted(pl, reverse = True) if i <= p] if len(poss) == 0 or s.solved: s.total += 1 return curList if abs(poss[0]-p) < 0.00001: s.solved = True # Solved it!!! s.total += 1 return curList + [poss[0]] ml,md = [], 10**8 for j in [s.solve(p-i, pl, [i]) for i in poss]: if abs(sum(j)-p)<md: ml,md = j, abs(sum(j)-p) s.total += 1 return ml + curList priceList = [5.8, 4.2, 3.55, 3.35, 2.75, 2.15] appetizers = ['Sampler Plate', 'Mozzarella Sticks', \ 'Hot wings', 'Side salad', 'French Fries', 'Mixed Fruit'] menu = zip(priceList, appetizers) sol = Solver() q = sol.solve(15.05, priceList) print 'Total time it runned: ', sol.total print '-'*30 order = [(m, q.count(m[0])) for m in menu if m[0] in q] for o in order: print '%d x %s \t\t\t (%.2f)' % (o[1],o[0][1],o[0][0]) print '-'*30 ts = 'Total: %.2f' % sum(q) print ' '*(30-len(ts)-1),ts
在python中。
我在使用"全局变量"时遇到了一些问题,因此我将函数作为对象的方法。它是递归的,它会在漫画中自问29次问题,并在第一次成功比赛后停止
Total time it runned: 29 ------------------------------ 1 x Sampler Plate (5.80) 2 x Hot wings (3.55) 1 x Mixed Fruit (2.15) ------------------------------ Total: 15.05
输出:
如果我们需要优化的算法,则最好按降序尝试价格。这样一来,我们可以先用尽所有剩余的金额,然后查看如何填充剩余的金额。
此外,我们可以使用数学方法来计算每次开始的每种食品的最大数量,因此我们不会尝试超过15.05美元目标的组合。
public class NPComplete { private static final int[] FOOD = { 580, 420, 355, 335, 275, 215 }; private static int tries; public static void main(String[] ignore) { tries = 0; addFood(1505, "", 0); System.out.println("Combinations tried: " + tries); } private static void addFood(int goal, String result, int index) { // If no more food to add, see if this is a solution if (index >= FOOD.length) { tries++; if (goal == 0) System.out.println(tries + " tries: " + result.substring(3)); return; } // Try all possible quantities of this food // If this is the last food item, only try the max quantity int qty = goal / FOOD[index]; do { addFood(goal - qty * FOOD[index], result + " + " + qty + " * " + FOOD[index], index + 1); } while (index < FOOD.length - 1 && --qty >= 0); } }
该算法只需尝试88种组合即可获得完整答案,这看起来是迄今为止发布的最低答案:
9 tries: 1 * 580 + 0 * 420 + 2 * 355 + 0 * 335 + 0 * 275 + 1 * 215 88 tries: 0 * 580 + 0 * 420 + 0 * 355 + 0 * 335 + 0 * 275 + 7 * 215 Combinations tried: 88
段落数量不匹配