下面是我的java实现:https://github.com/zouzhile/interview/blob/master/src/com/interview/algorithms/tree/BoggleSolver.java

Trie构建耗时0小时0分1秒532毫秒 单词搜索花了0小时0分0秒92毫秒

eel eeler eely eer eke eker eld eleut elk ell 
elle epee epihippus ere erept err error erupt eurus eye 
eyer eyey hip hipe hiper hippish hipple hippus his hish 
hiss hist hler hsi ihi iphis isis issue issuer ist 
isurus kee keek keeker keel keeler keep keeper keld kele 
kelek kelep kelk kell kelly kelp kelper kep kepi kept 
ker kerel kern keup keuper key kyl kyle lee leek 
leeky leep leer lek leo leper leptus lepus ler leu 
ley lleu lue lull luller lulu lunn lunt lunule luo 
lupe lupis lupulus lupus lur lure lurer lush lushly lust 
lustrous lut lye nul null nun nupe nurture nurturer nut 
oer ore ort ouphish our oust out outpeep outpeer outpipe 
outpull outpush output outre outrun outrush outspell outspue outspurn outspurt 
outstrut outstunt outsulk outturn outusure oyer pee peek peel peele 
peeler peeoy peep peeper peepeye peer pele peleus pell peller 
pelu pep peplus pepper pepperer pepsis per pern pert pertussis 
peru perule perun peul phi pip pipe piper pipi pipistrel 
pipistrelle pipistrellus pipper pish piss pist plup plus plush ply 
plyer psi pst puerer pul pule puler pulk pull puller 
pulley pullus pulp pulper pulu puly pun punt pup puppis 
pur pure puree purely purer purr purre purree purrel purrer 
puru purupuru pus push puss pustule put putt puture ree 
reek reeker reeky reel reeler reeper rel rely reoutput rep 
repel repeller repipe reply repp reps reree rereel rerun reuel 
roe roer roey roue rouelle roun roup rouper roust rout 
roy rue ruelle ruer rule ruler rull ruller run runt 
rupee rupert rupture ruru rus rush russ rust rustre rut 
shi shih ship shipper shish shlu sip sipe siper sipper 
sis sish sisi siss sissu sist sistrurus speel speer spelk 
spell speller splurt spun spur spurn spurrer spurt sput ssi 
ssu stre stree streek streel streeler streep streke streperous strepsis 
strey stroup stroy stroyer strue strunt strut stu stue stull 
stuller stun stunt stupe stupeous stupp sturnus sturt stuss stut 
sue suer suerre suld sulk sulker sulky sull sully sulu 
sun sunn sunt sunup sup supe super superoutput supper supple 
supplely supply sur sure surely surrey sus susi susu susurr 
susurrous susurrus sutu suture suu tree treey trek trekker trey 
troupe trouper trout troy true truer trull truller truly trun 
trush truss trust tshi tst tsun tsutsutsi tue tule tulle 
tulu tun tunu tup tupek tupi tur turn turnup turr 
turus tush tussis tussur tut tuts tutu tutulus ule ull 
uller ulu ululu unreel unrule unruly unrun unrust untrue untruly 
untruss untrust unturn unurn upper upperer uppish uppishly uppull uppush 
upspurt upsun upsup uptree uptruss upturn ure urn uro uru 
urus urushi ush ust usun usure usurer utu yee yeel 
yeld yelk yell yeller yelp yelper yeo yep yer yere 
yern yoe yor yore you youl youp your yourn yoy 

注意: 我在这个线程的开头使用了字典和字符矩阵。代码在我的MacBookPro上运行，下面是关于这台机器的一些信息。

型号:MacBook Pro 型号标识符:MacBookPro8,1 处理器名称:Intel Core i5 处理器速度:2.3 GHz 处理器数量:1 总核数:2 L2缓存(每核):256kb L3 Cache: 3mb 内存:4gb 引导ROM版本:MBP81.0047.B0E SMC版本(系统):1.68f96

你可以把这个问题分成两部分:

某种搜索算法可以在网格中列举出可能的字符串。 一种测试字符串是否是有效单词的方法。

理想情况下，(2)还应该包括一种测试字符串是否是有效单词前缀的方法——这将允许您精简搜索并节省大量时间。

亚当·罗森菲尔德(Adam Rosenfield)的Trie是(2)的一个解决方案。它很优雅，可能是算法专家的首选，但有了现代语言和现代计算机，我们可能会更懒一点。此外，正如Kent所建议的，我们可以通过丢弃网格中没有字母的单词来减少字典的大小。这是一些蟒蛇:

def make_lookups(grid, fn='dict.txt'):
    # Make set of valid characters.
    chars = set()
    for word in grid:
        chars.update(word)

    words = set(x.strip() for x in open(fn) if set(x.strip()) <= chars)
    prefixes = set()
    for w in words:
        for i in range(len(w)+1):
            prefixes.add(w[:i])

    return words, prefixes

哇;常数时间前缀测试。加载你链接的字典需要几秒钟，但只有几秒钟:-)(注意words <= prefixes)

现在，对于第(1)部分，我倾向于用图表来思考。所以我将创建一个像这样的字典:

graph = { (x, y):set([(x0,y0), (x1,y1), (x2,y2)]), }

例如，graph[(x, y)]是你从位置(x, y)可以到达的坐标集。我还将添加一个虚拟节点None，它将连接到所有东西。

构建它有点笨拙，因为有8个可能的位置，你必须做边界检查。下面是一些相应笨拙的python代码:

def make_graph(grid):
    root = None
    graph = { root:set() }
    chardict = { root:'' }

    for i, row in enumerate(grid):
        for j, char in enumerate(row):
            chardict[(i, j)] = char
            node = (i, j)
            children = set()
            graph[node] = children
            graph[root].add(node)
            add_children(node, children, grid)

    return graph, chardict

def add_children(node, children, grid):
    x0, y0 = node
    for i in [-1,0,1]:
        x = x0 + i
        if not (0 <= x < len(grid)):
            continue
        for j in [-1,0,1]:
            y = y0 + j
            if not (0 <= y < len(grid[0])) or (i == j == 0):
                continue

            children.add((x,y))

这段代码还建立了一个字典映射(x,y)到相应的字符。这让我把一个位置列表转换成一个单词:

def to_word(chardict, pos_list):
    return ''.join(chardict[x] for x in pos_list)

最后，我们进行深度优先搜索。基本程序是:

搜索到达一个特定的节点。 检查到目前为止的路径是否可能是单词的一部分。如果不是，就不要进一步探索这个分支。 检查到目前为止的路径是否是一个单词。如果是，则添加到结果列表中。 探索迄今为止所有孩子未走的路。

Python:

def find_words(graph, chardict, position, prefix, results, words, prefixes):
    """ Arguments:
      graph :: mapping (x,y) to set of reachable positions
      chardict :: mapping (x,y) to character
      position :: current position (x,y) -- equals prefix[-1]
      prefix :: list of positions in current string
      results :: set of words found
      words :: set of valid words in the dictionary
      prefixes :: set of valid words or prefixes thereof
    """
    word = to_word(chardict, prefix)

    if word not in prefixes:
        return

    if word in words:
        results.add(word)

    for child in graph[position]:
        if child not in prefix:
            find_words(graph, chardict, child, prefix+[child], results, words, prefixes)

运行代码如下:

grid = ['fxie', 'amlo', 'ewbx', 'astu']
g, c = make_graph(grid)
w, p = make_lookups(grid)
res = set()
find_words(g, c, None, [], res, w, p)

检查保留区，看看答案。下面是为你的例子找到的单词列表，按大小排序:

 ['a', 'b', 'e', 'f', 'i', 'l', 'm', 'o', 's', 't',
 'u', 'w', 'x', 'ae', 'am', 'as', 'aw', 'ax', 'bo',
 'bu', 'ea', 'el', 'em', 'es', 'fa', 'ie', 'io', 'li',
 'lo', 'ma', 'me', 'mi', 'oe', 'ox', 'sa', 'se', 'st',
 'tu', 'ut', 'wa', 'we', 'xi', 'aes', 'ame', 'ami',
 'ase', 'ast', 'awa', 'awe', 'awl', 'blo', 'but', 'elb',
 'elm', 'fae', 'fam', 'lei', 'lie', 'lim', 'lob', 'lox',
 'mae', 'maw', 'mew', 'mil', 'mix', 'oil', 'olm', 'saw',
 'sea', 'sew', 'swa', 'tub', 'tux', 'twa', 'wae', 'was',
 'wax', 'wem', 'ambo', 'amil', 'amli', 'asem', 'axil',
 'axle', 'bleo', 'boil', 'bole', 'east', 'fame', 'limb',
 'lime', 'mesa', 'mewl', 'mile', 'milo', 'oime', 'sawt',
 'seam', 'seax', 'semi', 'stub', 'swam', 'twae', 'twas',
 'wame', 'wase', 'wast', 'weam', 'west', 'amble', 'awest',
 'axile', 'embox', 'limbo', 'limes', 'swami', 'embole',
 'famble', 'semble', 'wamble']

代码需要(字面上的)几秒钟来加载字典，但其余的在我的机器上是立即完成的。

我用c语言解决了这个问题。在我的机器上运行大约需要48毫秒(其中98%的时间花在从磁盘加载字典和创建trie上)。字典是/usr/share/dict/american-english，有62886个单词。

源代码

对于字典加速，有一个通用的转换/过程可以大大减少提前的字典比较。

鉴于上面的网格只包含16个字符，其中一些字符是重复的，您可以通过简单地过滤掉具有不可获取字符的条目来大大减少字典中的总键数。

我认为这是明显的优化，但看到没有人这么做，我就提出来了。

在输入过程中，它将我的字典从20万个键减少到只有2000个键。这至少减少了内存开销，并且这肯定会映射到某个地方的速度增加，因为内存不是无限快的。

Perl实现

我的实现有点头重脚轻，因为我重视能够知道每个提取的字符串的确切路径，而不仅仅是其中的有效性。

我也有一些适应在那里，理论上允许一个网格中有洞的功能，网格有不同大小的线(假设你得到了正确的输入，它以某种方式对齐)。

早期筛选器是我的应用程序中最重要的瓶颈，正如之前怀疑的那样，注释掉了一行从1.5s膨胀到7.5s的代码。

在执行时，它似乎认为所有的个位数都在他们自己的有效单词上，但我很确定这是由于字典文件的工作方式。

它有点臃肿，但至少我重用了cpan中的Tree::Trie

其中有些部分是受到现有实现的启发，有些是我已经想到的。

建设性的批评和改进的方法欢迎(/我注意到他从来没有在CPAN上搜索过一个拼字游戏解决器，但这更有趣)

新标准更新

#!/usr/bin/perl 

use strict;
use warnings;

{

  # this package manages a given path through the grid.
  # Its an array of matrix-nodes in-order with
  # Convenience functions for pretty-printing the paths
  # and for extending paths as new paths.

  # Usage:
  # my $p = Prefix->new(path=>[ $startnode ]);
  # my $c = $p->child( $extensionNode );
  # print $c->current_word ;

  package Prefix;
  use Moose;

  has path => (
      isa     => 'ArrayRef[MatrixNode]',
      is      => 'rw',
      default => sub { [] },
  );
  has current_word => (
      isa        => 'Str',
      is         => 'rw',
      lazy_build => 1,
  );

  # Create a clone of this object
  # with a longer path

  # $o->child( $successive-node-on-graph );

  sub child {
      my $self    = shift;
      my $newNode = shift;
      my $f       = Prefix->new();

      # Have to do this manually or other recorded paths get modified
      push @{ $f->{path} }, @{ $self->{path} }, $newNode;
      return $f;
  }

  # Traverses $o->path left-to-right to get the string it represents.

  sub _build_current_word {
      my $self = shift;
      return join q{}, map { $_->{value} } @{ $self->{path} };
  }

  # Returns  the rightmost node on this path

  sub tail {
      my $self = shift;
      return $self->{path}->[-1];
  }

  # pretty-format $o->path

  sub pp_path {
      my $self = shift;
      my @path =
        map { '[' . $_->{x_position} . ',' . $_->{y_position} . ']' }
        @{ $self->{path} };
      return "[" . join( ",", @path ) . "]";
  }

  # pretty-format $o
  sub pp {
      my $self = shift;
      return $self->current_word . ' => ' . $self->pp_path;
  }

  __PACKAGE__->meta->make_immutable;
}

{

  # Basic package for tracking node data
  # without having to look on the grid.
  # I could have just used an array or a hash, but that got ugly.

# Once the matrix is up and running it doesn't really care so much about rows/columns,
# Its just a sea of points and each point has adjacent points.
# Relative positioning is only really useful to map it back to userspace

  package MatrixNode;
  use Moose;

  has x_position => ( isa => 'Int', is => 'rw', required => 1 );
  has y_position => ( isa => 'Int', is => 'rw', required => 1 );
  has value      => ( isa => 'Str', is => 'rw', required => 1 );
  has siblings   => (
      isa     => 'ArrayRef[MatrixNode]',
      is      => 'rw',
      default => sub { [] }
  );

# Its not implicitly uni-directional joins. It would be more effient in therory
# to make the link go both ways at the same time, but thats too hard to program around.
# and besides, this isn't slow enough to bother caring about.

  sub add_sibling {
      my $self    = shift;
      my $sibling = shift;
      push @{ $self->siblings }, $sibling;
  }

  # Convenience method to derive a path starting at this node

  sub to_path {
      my $self = shift;
      return Prefix->new( path => [$self] );
  }
  __PACKAGE__->meta->make_immutable;

}

{

  package Matrix;
  use Moose;

  has rows => (
      isa     => 'ArrayRef',
      is      => 'rw',
      default => sub { [] },
  );

  has regex => (
      isa        => 'Regexp',
      is         => 'rw',
      lazy_build => 1,
  );

  has cells => (
      isa        => 'ArrayRef',
      is         => 'rw',
      lazy_build => 1,
  );

  sub add_row {
      my $self = shift;
      push @{ $self->rows }, [@_];
  }

  # Most of these functions from here down are just builder functions,
  # or utilities to help build things.
  # Some just broken out to make it easier for me to process.
  # All thats really useful is add_row
  # The rest will generally be computed, stored, and ready to go
  # from ->cells by the time either ->cells or ->regex are called.

  # traverse all cells and make a regex that covers them.
  sub _build_regex {
      my $self  = shift;
      my $chars = q{};
      for my $cell ( @{ $self->cells } ) {
          $chars .= $cell->value();
      }
      $chars = "[^$chars]";
      return qr/$chars/i;
  }

  # convert a plain cell ( ie: [x][y] = 0 )
  # to an intelligent cell ie: [x][y] = object( x, y )
  # we only really keep them in this format temporarily
  # so we can go through and tie in neighbouring information.
  # after the neigbouring is done, the grid should be considered inoperative.

  sub _convert {
      my $self = shift;
      my $x    = shift;
      my $y    = shift;
      my $v    = $self->_read( $x, $y );
      my $n    = MatrixNode->new(
          x_position => $x,
          y_position => $y,
          value      => $v,
      );
      $self->_write( $x, $y, $n );
      return $n;
  }

# go through the rows/collums presently available and freeze them into objects.

  sub _build_cells {
      my $self = shift;
      my @out  = ();
      my @rows = @{ $self->{rows} };
      for my $x ( 0 .. $#rows ) {
          next unless defined $self->{rows}->[$x];
          my @col = @{ $self->{rows}->[$x] };
          for my $y ( 0 .. $#col ) {
              next unless defined $self->{rows}->[$x]->[$y];
              push @out, $self->_convert( $x, $y );
          }
      }
      for my $c (@out) {
          for my $n ( $self->_neighbours( $c->x_position, $c->y_position ) ) {
              $c->add_sibling( $self->{rows}->[ $n->[0] ]->[ $n->[1] ] );
          }
      }
      return \@out;
  }

  # given x,y , return array of points that refer to valid neighbours.
  sub _neighbours {
      my $self = shift;
      my $x    = shift;
      my $y    = shift;
      my @out  = ();
      for my $sx ( -1, 0, 1 ) {
          next if $sx + $x < 0;
          next if not defined $self->{rows}->[ $sx + $x ];
          for my $sy ( -1, 0, 1 ) {
              next if $sx == 0 && $sy == 0;
              next if $sy + $y < 0;
              next if not defined $self->{rows}->[ $sx + $x ]->[ $sy + $y ];
              push @out, [ $sx + $x, $sy + $y ];
          }
      }
      return @out;
  }

  sub _has_row {
      my $self = shift;
      my $x    = shift;
      return defined $self->{rows}->[$x];
  }

  sub _has_cell {
      my $self = shift;
      my $x    = shift;
      my $y    = shift;
      return defined $self->{rows}->[$x]->[$y];
  }

  sub _read {
      my $self = shift;
      my $x    = shift;
      my $y    = shift;
      return $self->{rows}->[$x]->[$y];
  }

  sub _write {
      my $self = shift;
      my $x    = shift;
      my $y    = shift;
      my $v    = shift;
      $self->{rows}->[$x]->[$y] = $v;
      return $v;
  }

  __PACKAGE__->meta->make_immutable;
}

use Tree::Trie;

sub readDict {
  my $fn = shift;
  my $re = shift;
  my $d  = Tree::Trie->new();

  # Dictionary Loading
  open my $fh, '<', $fn;
  while ( my $line = <$fh> ) {
      chomp($line);

 # Commenting the next line makes it go from 1.5 seconds to 7.5 seconds. EPIC.
      next if $line =~ $re;    # Early Filter
      $d->add( uc($line) );
  }
  return $d;
}

sub traverseGraph {
  my $d     = shift;
  my $m     = shift;
  my $min   = shift;
  my $max   = shift;
  my @words = ();

  # Inject all grid nodes into the processing queue.

  my @queue =
    grep { $d->lookup( $_->current_word ) }
    map  { $_->to_path } @{ $m->cells };

  while (@queue) {
      my $item = shift @queue;

      # put the dictionary into "exact match" mode.

      $d->deepsearch('exact');

      my $cword = $item->current_word;
      my $l     = length($cword);

      if ( $l >= $min && $d->lookup($cword) ) {
          push @words,
            $item;    # push current path into "words" if it exactly matches.
      }
      next if $l > $max;

      # put the dictionary into "is-a-prefix" mode.
      $d->deepsearch('boolean');

    siblingloop: foreach my $sibling ( @{ $item->tail->siblings } ) {
          foreach my $visited ( @{ $item->{path} } ) {
              next siblingloop if $sibling == $visited;
          }

          # given path y , iterate for all its end points
          my $subpath = $item->child($sibling);

          # create a new path for each end-point
          if ( $d->lookup( $subpath->current_word ) ) {

             # if the new path is a prefix, add it to the bottom of the queue.
              push @queue, $subpath;
          }
      }
  }
  return \@words;
}

sub setup_predetermined { 
  my $m = shift; 
  my $gameNo = shift;
  if( $gameNo == 0 ){
      $m->add_row(qw( F X I E ));
      $m->add_row(qw( A M L O ));
      $m->add_row(qw( E W B X ));
      $m->add_row(qw( A S T U ));
      return $m;
  }
  if( $gameNo == 1 ){
      $m->add_row(qw( D G H I ));
      $m->add_row(qw( K L P S ));
      $m->add_row(qw( Y E U T ));
      $m->add_row(qw( E O R N ));
      return $m;
  }
}
sub setup_random { 
  my $m = shift; 
  my $seed = shift;
  srand $seed;
  my @letters = 'A' .. 'Z' ; 
  for( 1 .. 4 ){ 
      my @r = ();
      for( 1 .. 4 ){
          push @r , $letters[int(rand(25))];
      }
      $m->add_row( @r );
  }
}

# Here is where the real work starts.

my $m = Matrix->new();
setup_predetermined( $m, 0 );
#setup_random( $m, 5 );

my $d = readDict( 'dict.txt', $m->regex );
my $c = scalar @{ $m->cells };    # get the max, as per spec

print join ",\n", map { $_->pp } @{
  traverseGraph( $d, $m, 3, $c ) ;
};

Arch/执行信息进行比较:

model name      : Intel(R) Core(TM)2 Duo CPU     T9300  @ 2.50GHz
cache size      : 6144 KB
Memory usage summary: heap total: 77057577, heap peak: 11446200, stack peak: 26448
       total calls   total memory   failed calls
 malloc|     947212       68763684              0
realloc|      11191        1045641              0  (nomove:9063, dec:4731, free:0)
 calloc|     121001        7248252              0
   free|     973159       65854762

Histogram for block sizes:
  0-15         392633  36% ==================================================
 16-31          43530   4% =====
 32-47          50048   4% ======
 48-63          70701   6% =========
 64-79          18831   1% ==
 80-95          19271   1% ==
 96-111        238398  22% ==============================
112-127          3007  <1% 
128-143        236727  21% ==============================

关于正则表达式优化的更多嘟囔

我使用的正则表达式优化对于多解字典是无用的，而对于多解字典，您将需要一个完整的字典，而不是一个预先修整过的字典。

然而，也就是说，对于一次性解决，它真的很快。(Perl正则表达式是在C!:))

以下是一些不同的代码添加:

sub readDict_nofilter {
  my $fn = shift;
  my $re = shift;
  my $d  = Tree::Trie->new();

  # Dictionary Loading
  open my $fh, '<', $fn;
  while ( my $line = <$fh> ) {
      chomp($line);
      $d->add( uc($line) );
  }
  return $d;
}

sub benchmark_io { 
  use Benchmark qw( cmpthese :hireswallclock );
   # generate a random 16 character string 
   # to simulate there being an input grid. 
  my $regexen = sub { 
      my @letters = 'A' .. 'Z' ; 
      my @lo = ();
      for( 1..16 ){ 
          push @lo , $_ ; 
      }
      my $c  = join '', @lo;
      $c = "[^$c]";
      return qr/$c/i;
  };
  cmpthese( 200 , { 
      filtered => sub { 
          readDict('dict.txt', $regexen->() );
      }, 
      unfiltered => sub {
          readDict_nofilter('dict.txt');
      }
  });
}

           s/iter unfiltered   filtered
unfiltered   8.16         --       -94%
filtered    0.464      1658%         --

Ps: 8.16 * 200 = 27分钟。

我已经在OCaml中实现了一个解决方案。它将字典预编译为trie，并使用双字母序列频率来消除单词中永远不会出现的边，以进一步加快处理速度。

它在0.35ms内解决了示例板的问题(额外的6ms启动时间主要与将trie加载到内存有关)。

找到的解决方案:

["swami"; "emile"; "limbs"; "limbo"; "limes"; "amble"; "tubs"; "stub";
 "swam"; "semi"; "seam"; "awes"; "buts"; "bole"; "boil"; "west"; "east";
 "emil"; "lobs"; "limb"; "lime"; "lima"; "mesa"; "mews"; "mewl"; "maws";
 "milo"; "mile"; "awes"; "amie"; "axle"; "elma"; "fame"; "ubs"; "tux"; "tub";
 "twa"; "twa"; "stu"; "saw"; "sea"; "sew"; "sea"; "awe"; "awl"; "but"; "btu";
 "box"; "bmw"; "was"; "wax"; "oil"; "lox"; "lob"; "leo"; "lei"; "lie"; "mes";
 "mew"; "mae"; "maw"; "max"; "mil"; "mix"; "awe"; "awl"; "elm"; "eli"; "fax"]

你的搜索算法是否会随着搜索的继续而不断减少单词列表?

例如，在上面的搜索中，你的单词只能以13个字母开头(有效地减少了一半的开头字母)。

当你添加更多的字母排列时，它会进一步减少可用的单词集，减少必要的搜索。

我会从这里开始。