较高人工智能的人机博弈程序实现(多个算法结合)含C++源码
本文由恋花蝶最初发表于http://blog.youkuaiyun.com/lanphaday 上,您可以转载、引用、打印和分发等,但必须保留本文完整和包含本声明,否则必究责任。
到昨天晚上,Topcoder Marathon Match 6结束了,我取得了第18名的成绩,已经是自己参加Marathon四次以来的最好名次啦,高兴ing。因为这次的题目比较偏:写一个人工智能程序和服务器端的程序进行博弈。人机博弈是一门比较专的学科,大部分中国高手都不能快速的在比赛中学习和实现一些复杂的算法,以致成绩不太如意;我挟之前对这方面的了解,做得还算行,所以把代码公开出来,可以多一点中文方面的资料和源码给大家参考,我也感到非常荣幸。
比赛的题目请看这里:http://www.topcoder.com/longcontest/?module=ViewProblemStatement&rd=10118&pm=6759 主要的游戏规则也是在这里的,我就不在这里重复啦,主要讲讲我的代码用到了什么算法。麻将虽小,五脏俱全,主要应用的算法有主要变量搜索(PVS)、历史启发(HH)、杀手启发(KH)、Null Move和迭代深化(ID),可惜后来不够时间实现置换表(TT),不然可以多一个算法了。代码里还实现了时间控制策略,可以几乎用尽20秒的测试时间,为争取更好的着法提供了保证。还有值得一提的是棋盘表示,我使用了棋盘表、棋子位置表结合的方式来表示,后来发现加上空位表的话,可以加快不少走法生成和估值的速度。反正棋盘表示是一切的基础,一种好的表示方法可以带来很大的性能提升。对于代码,大家注意class SE里的search_move和pvs两个函数,上述的算法和策略都在那里。class MG是关于棋盘表示、走法生成和估值的,class KH和class HH分别是杀手启发和历史启发。Null Move是简单有效的算法,不过我的实现里是比较简单的那种,如果有兴趣,可以查询其它资料。
讲了这么多,应该说一下这份代码的计算能力:以6*6的棋盘为例,这份代码在VC6的release模式下编译运行可以在1秒内搜索并评估83万个叶子节点,计算层次在8-9层;如果用MiniMax算法不进行剪枝,只能搜索到3-4层(测试机器皆为超线程P4 3.0G+1G内存)。这就是算法的力量吧。另声明一下,本代码未作优化,不代表我不懂,只是没有时间,看的朋友请海涵了。
下面是代码,在VC和G++上皆可编译、执行
因为比赛期间写的,代码比较乱,但整体的风格还是可以的,复制到IDE上看可能会更好看些
#include
<
iostream
>
#include
<
cstdlib
>
#include
<
ctime
>
#include
<
cassert
>
#include
<
vector
>
#include
<
algorithm
>

using
namespace
std;
typedefunsigned
int
UINT;
typedefUINTMOVE;
const
int
INFINITY
=
100000000
;
const
int
MAX_DEPTH
=
16
;
const
UINTmax_board_size
=
256
;
const
UINTmax_stones_cnt
=
256
;
const
UINTempty
=
0
;
const
UINTmy_color
=
1
;
const
UINTsvr_color
=
2
;
#ifdefWIN32
const
clock_tall_time
=
19200
;
#else
const
clock_tall_time
=
19200000
;
#endif

const
UINTcheck_time_cnt
=
0x00001fff
;
#define
is_empty(x)(x==empty)

#define
opp_color(x)(x==my_color?svr_color:my_color)

int
leaf_cnt
=
0
;
class
MG
...
{
private :
UINTN_;
UINTboard_[max_board_size];
UINTstones_[max_stones_cnt];
private :
void extend(UINTpos,unsigned char * eht,unsigned char * est,UINT & area,UINT & round);
public :
MOVEmove_table[MAX_DEPTH][max_board_size];
UINTcurr_stones_cnt;
UINTcurr_board_size;
void set_N( int n) ... {
N_ = n;
curr_board_size = n * n;
curr_stones_cnt = 0 ;
memset(board_, 0 , sizeof (UINT) * max_board_size);
memset(stones_, 0 , sizeof (UINT) * max_stones_cnt);
} 
void make_move( int idx, int color) ... {
board_[idx] = color;
stones_[curr_stones_cnt ++ ] = idx;
} 
void unmake_move( int idx) ... {
board_[idx] = empty;
-- curr_stones_cnt;
} 
inline bool is_game_over() ... { return curr_stones_cnt == curr_board_size;}
UINTgen_move( int depth);
int evaluatoin( int color);
int evaluatoin_4_end( int color);
void print_board()
... {
int cnt = 0 ;
for (UINTi = 0 ;i < curr_board_size; ++ i)
... {
if (is_empty(board_[i]))
cout << " o " ;
else
cout << ((board_[i] == my_color) ? " @ " : " - " );
++ cnt;
if (cnt == N_)
... {
cnt = 0 ;
cout << endl;
}
}
} 
bool can_move(MOVEmove) ... { return is_empty(board_[move]);}
void remove_killers( int depth, int move_cnt,MOVE * killers, int killers_cnt)
... {
for ( int i = 0 ;i < killers_cnt; ++ i)
... {
MOVEm = killers[i];
for ( int j = 0 ;j < move_cnt; ++ j)
... {
if (move_table[depth][j] != m)
continue ;
for ( int k = j + 1 ;k < move_cnt; ++ k)
... {
move_table[depth][k - 1 ] = move_table[depth][k];
}
break ;
}
}
}
}
;
UINTMG::gen_move(
int
depth)
...
{
int cnt = 0 ;
for (UINTi = 0 ;i < curr_board_size; ++ i)
... {
if (is_empty(board_[i]))
move_table[depth][cnt ++ ] = i;
}
return cnt;
}

int
MG::evaluatoin(
int
color)
...
{
if (curr_stones_cnt + 1 == curr_board_size)
... {
for ( int i = 0 ;i < curr_board_size; ++ i)
... {
if (is_empty(board_[i]))
... {
board_[i] = color;
int value = - evaluatoin_4_end(opp_color(color));
board_[i] = empty;
return value;
}
}
}
++ leaf_cnt;
unsigned char extended_hash_table[max_board_size] = ... { 0 } ;
int my_score = 0 ,svr_score = 0 ;
for (UINTi = 0 ;i < curr_stones_cnt; ++ i)
... {
UINTpos = stones_[i];
if (extended_hash_table[pos])
continue ;
UINTarea = 0 ,round = 0 ;
unsigned char extended_space_table[max_board_size] = ... { 0 } ;
extend(pos,extended_hash_table,extended_space_table,area,round);
if (board_[pos] == my_color)
... {
my_score += area * area * round;
}
else 
... {
svr_score += area * area * round;
}
}
if (color == my_color)
return my_score - svr_score;
return svr_score - my_score;
}

int
MG::evaluatoin_4_end(
int
color)
...
{
++ leaf_cnt;
unsigned char extended_hash_table[max_board_size] = ... { 0 } ;
int my_score = 0 ,svr_score = 0 ;
for (UINTi = 0 ;i < curr_stones_cnt; ++ i)
... {
UINTpos = stones_[i];
if (extended_hash_table[pos])
continue ;
UINTarea = 0 ,round = 0 ;
unsigned char extended_space_table[max_board_size] = ... { 0 } ;
extend(pos,extended_hash_table,extended_space_table,area,round);
if (board_[pos] == my_color)
... {
my_score += area * area;
}
else 
... {
svr_score += area * area;
}
}
if (color == my_color)
return my_score - svr_score;
return svr_score - my_score;
}

void
MG::extend(UINTpos,unsigned
char
*
eht,unsigned
char
*
est,UINT
&
area,UINT
&
round)
...
{
const UINTround_cnt = 4 ;
int is [round_cnt] = ... { - N_, - 1 , 1 ,N_} ;
++ area;
eht[pos] = 1 ;
for (UINTi = 0 ;i < round_cnt; ++ i)
... {
int new_idx = pos + is [i];
if (new_idx < 0 || new_idx >= curr_board_size)
continue ;
if (i == 1 && pos % N_ == 0 )
continue ;
if (i == 2 && new_idx % N_ == 0 )
continue ;
if (is_empty(board_[new_idx]) && ( ! est[new_idx]))
... {
++ round;
est[new_idx] = 1 ;
continue ;
}
if (eht[new_idx])
continue ;
if (board_[new_idx] == board_[pos])
extend(new_idx,eht,est,area,round);
}
}

class
HH
...
{
private :
UINTboard_[ 2 ][max_board_size];
public :
void reset() ... {memset(board_, 0 , sizeof (UINT) * max_board_size);}
void update_value( int depth, int color,MOVEmove);
MOVEget_best(MOVE * move_list, int color, int cnt);
}
;
void
HH::update_value(
int
depth,
int
color,MOVEmove)
...
{
board_[color - 1 ][move] += ( 1 << depth);
}

MOVEHH::get_best(MOVE
*
move_list,
int
color,
int
cnt)
...
{
int real_color = color - 1 ;
MOVE * p = move_list;
int best = board_[real_color][ * move_list];
int best_idx = 0 ;
for ( int i = 1 ;i < cnt; ++ i)
... {
++ move_list;
if (board_[real_color][ * move_list] <= best)
continue ;
best = board_[real_color][ * move_list];
best_idx = i;
}
MOVEtmp = * p;
* p = p[best_idx];
p[best_idx] = tmp;
return * p;
}

struct
KH_item
...
{
MOVEmove;
int cnt;
}
;
class
less_than
...
{
public :
inline bool operator ()( const KH_item & lhs, const KH_item & rhs)
... {
return lhs.cnt < rhs.cnt;
}
}
;
const
int
max_kh_item_cnt
=
4
;
class
KH
...
{
private :
KH_itemKH_table[MAX_DEPTH][max_kh_item_cnt];
int cnt_table[MAX_DEPTH];
public :
void add_to_kh(MOVEmove, int depth)
... {
int cnt_mini_idx = 0 ;
int cnt_mini = KH_table[depth][ 0 ].cnt;
int i = 0 ;
for (i = 0 ;i < cnt_table[depth]; ++ i)
... {
KH_item & tmp = KH_table[depth][i];
if (tmp.move == move)
... {
++ tmp.cnt;
return ;
}
if (tmp.cnt < cnt_mini)
... {
cnt_mini_idx = i;
cnt_mini = tmp.cnt;
}
}
if (i < max_kh_item_cnt)
... {
KH_table[depth][i].move = move;
++ (cnt_table[depth]);
}
else 
... {
KH_item & tmp = KH_table[depth][cnt_mini_idx];
tmp.move = move;
tmp.cnt = 1 ;
}
}
int get_killers(MOVE * killers, int depth)
... {
sort < KH_item *> (KH_table[depth],KH_table[depth] + cnt_table[depth],less_than());
int i = 0 ;
for (i = 0 ;i < cnt_table[depth]; ++ i)
... {
killers[i] = KH_table[depth][i].move;
}
return i;
}
void reset()
... {
memset(cnt_table, 0 , sizeof ( int ) * MAX_DEPTH);
memset(KH_table, 0 , sizeof (KH_item) * MAX_DEPTH * max_kh_item_cnt);
}
}
;
class
SE
...
{
private :
MGmg;
HHhh;
KHkh;
int N_;
int best_move;
int max_depth_;
public :
void print_board()
... {
mg.print_board();
}
void set_N( int N)
... {
N_ = N;
used_time = 0 ;
best_move = 0xffff ;
mg.set_N(N);
}
vector < int > get_best_move()
... {
int row = best_move / N_;
int col = best_move % N_;
vector < int > move;
move.push_back(row);
move.push_back(col);
return move;
}
void do_move( int row, int col, int color)
... {
mg.make_move(row * N_ + col,color);
}
void make_sure_best_move_first(MOVE * moves, int cnt,MOVEbest_move);
vector < int > search_move( int max_depth);
int pvs( int , int , int , int , int );
private :
clock_tbgn_time;
clock_tused_time;
clock_tcurr_time_limit;
}
;
void
SE::make_sure_best_move_first(MOVE
*
moves,
int
cnt,MOVEbest_move)
...
{
for ( int i = 0 ;i < cnt; ++ i)
... {
if (moves[i] == best_move)
... {
moves[i] = moves[ 0 ];
moves[ 0 ] = best_move;
}
}
}

vector
<
int
>
SE::search_move(
int
max_depth)
...
{
leaf_cnt = 1 ;
bgn_time = clock(); // ³õʼʱ¼ä
// ¼ÆËã±¾´ÎʱÏÞ
UINTleave_space_cnt = mg.curr_board_size - mg.curr_stones_cnt;
if (leave_space_cnt >= 2 )
leave_space_cnt /= 2 ;
curr_time_limit = (all_time - used_time) / leave_space_cnt;
if (curr_time_limit > all_time || curr_time_limit < 0 )
... {
curr_time_limit = 1 ;
} 
if (leave_space_cnt < mg.curr_board_size / 3 )
curr_time_limit = (( double )curr_time_limit) * ( 1.4 );
else if (leave_space_cnt < mg.curr_board_size / 2 )
curr_time_limit = (( double )curr_time_limit) * ( 1.3 );
if (N_ > 12 )
curr_time_limit = (( double )curr_time_limit) * ( 0.9 );
hh.reset();
kh.reset();
int md = 0 ;
int backup_max_depth = max_depth;
while (md < max_depth)
... {
++ md;
max_depth_ = md;
pvs(md,my_color, 0 , - INFINITY,INFINITY);
if (max_depth >= backup_max_depth)
... {
// »¹ÓÐʱ¼ä£¿
if (clock() - bgn_time < curr_time_limit)
... {
// ²»»á¶ÑÕ»Òç³ö£¿ÔÙËã¶àÒ»²ã
if (max_depth < MAX_DEPTH - 1 )
++ max_depth;
}
}
if (clock() - bgn_time >= curr_time_limit)
... {
break ;
}
}
clock_tcurr_used = clock() - bgn_time;
used_time += curr_used; // Ôö¼ÓÓõôµÄʱ¼ä
return get_best_move();
}

int
SE::pvs(
int
depth,
int
color,
int
nullmove,
int
alpha,
int
beta)
...
{
if (mg.is_game_over())
return mg.evaluatoin_4_end(color);
if (depth <= 0 )
return mg.evaluatoin(color);
if ((leaf_cnt & check_time_cnt) == 0 ) // ¼ì²âÊÇ·ñ³¬Ê± 
... {
if (clock() - bgn_time >= curr_time_limit)
return mg.evaluatoin(color);
} 
// NullMove
if (depth < max_depth_ && nullmove == 0 )
... {
int value = - pvs(depth - 2 ,opp_color(color), 1 , - alpha - 1 , - alpha);
if (value >= beta)
... {
return value;
}
} 
// killermove
int best;
MOVEbm = 0xffff ;
MOVEkillers[max_kh_item_cnt];
int killers_cnt = kh.get_killers(killers,depth);
if (killers_cnt > 0 && depth == max_depth_)
make_sure_best_move_first(killers,killers_cnt,best_move);
for ( int k = 0 ;k < killers_cnt; ++ k)
... {
MOVEm = killers[k];
if ( ! mg.can_move(m))
continue ;
mg.make_move(m,color);
best = - pvs(depth - 1 ,opp_color(color), 0 , - alpha - 1 , - alpha);
if (best >= beta)
... {
if (depth == max_depth_)
best_move = m;
kh.add_to_kh(m,depth);
hh.update_value(depth,color,m);
mg.unmake_move(m);
return best;
}
else if (best > alpha)
... {
alpha = best;
bm = m;
}
mg.unmake_move(m);
if ((leaf_cnt & check_time_cnt) == 0 ) // ¼ì²âÊÇ·ñ³¬Ê± 
... {
if (clock() - bgn_time >= curr_time_limit)
break ;
}
} 
// PVS
int move_cnt = mg.gen_move(depth);
if (depth == max_depth_)
make_sure_best_move_first(mg.move_table[depth],move_cnt,best_move);
if (killers_cnt == 0 || bm == 0xffff ) // bm==0xffff±íʾkillersÎÞЧ£¡ 
... {
if (depth == max_depth_)
bm = mg.move_table[depth][ 0 ];
else
bm = hh.get_best(mg.move_table[depth],color,move_cnt);
mg.make_move(bm,color);
best = - pvs(depth - 1 ,opp_color(color), 0 , - beta, - alpha);
mg.unmake_move(bm);
}
else 
... {
// removekillersfrommove_table
if (killers_cnt > 0 )
mg.remove_killers(depth,move_cnt,killers,killers_cnt);
MOVEbm_;
if (depth == max_depth_)
bm_ = mg.move_table[depth][ 0 ];
else
bm_ = hh.get_best(mg.move_table[depth],color,move_cnt);
mg.make_move(bm_,color);
int best_ = - pvs(depth - 1 ,opp_color(color), 0 , - beta, - alpha);
if (best_ > best)
... {
best = best_;
bm = bm_;
}
mg.unmake_move(bm_);
}
for ( int i = 1 ;i < move_cnt; ++ i)
... {
if (best >= beta)
break ;
if (best > alpha)
alpha = best;
if ((leaf_cnt & check_time_cnt) == 0 ) // ¼ì²âÊÇ·ñ³¬Ê± 
... {
if (clock() - bgn_time >= curr_time_limit)
break ;
} 
MOVEm = hh.get_best(mg.move_table[depth] + i,color,move_cnt - i);
mg.make_move(m,color);
int value = - pvs(depth - 1 ,opp_color(color), 0 , - alpha - 1 , - alpha);
if (value > alpha && value < beta)
... {
best = - pvs(depth - 1 ,opp_color(color), 0 , - beta, - value);
bm = m;
}
else if (value > best)
... {
best = value;
bm = m;
}
mg.unmake_move(m);
}
if (depth == max_depth_)
best_move = bm;
if (best >= alpha)
... {
kh.add_to_kh(bm,depth);
hh.update_value(depth,color,bm);
}
return best;
}

class
PseudoTonga
...
{
public :
vector < int > move( int row, int col);
vector < int > init( int N, int row, int col);
private :
int N_;
SEse;
void do_move( int row, int col, int color);
}
;
vector
<
int
>
PseudoTonga::init(
int
N,
int
row,
int
col)
...
{
N_ = N;
se.set_N(N);
int r = 0 ,c = 0 ;
if (row >= 0 || col >= 0 )
... {
return move(row,col);
} 
vector < int > move;
r = c = N / 2 ;
do_move(r,c,my_color);
move.push_back(r);
move.push_back(c);
cout << " player:row= " << move[ 0 ] << " , col= " << move[ 1 ] << " ; " ;
return move;
}

vector
<
int
>
PseudoTonga::move(
int
row,
int
col)
...
{
do_move(row,col,svr_color);
cout << " server:row= " << row << " , col= " << col << " ; " ;
vector < int > move;
int d = 3 ;
move = se.search_move(d);
do_move(move[ 0 ],move[ 1 ],my_color);
cout << " player:row= " << move[ 0 ] << " , col= " << move[ 1 ] << " ; " ;
cout << " leafcountis " << leaf_cnt << endl;
return move;
}

void
PseudoTonga::do_move(
int
row,
int
col,
int
color)
...
{
se.do_move(row,col,color);
}
int main()
{
PseudoTonga pt;
pt.init(6, 2, 2);
pt.move(2,4);
return 0;
}
本文介绍了一个较高人工智能的人机博弈程序实现案例,通过多种算法结合,如PVS、HH、KH等,显著提升了搜索效率和博弈水平。代码在特定配置下每秒能评估83万个叶子节点,达到了8-9层的计算深度。
2万+

被折叠的 条评论
为什么被折叠?



