Author: Andrew Williams
Date: 02:00:27 04/14/04
Go up one level in this thread
On April 14, 2004 at 04:57:01, Dann Corbit wrote:
>On April 14, 2004 at 04:38:01, Dann Corbit wrote:
>
>>On April 14, 2004 at 04:26:10, Richard Pijl wrote:
>>
>>>On April 14, 2004 at 03:49:06, Dann Corbit wrote:
>>>
>>>>On April 14, 2004 at 03:46:08, Richard Pijl wrote:
>>>>
>>>>>On April 14, 2004 at 03:30:22, Dann Corbit wrote:
>>>>>
>>>>>>I decided to toss an MTD(f) search into TSCP, and I've got something wrong, but
>>>>>>I can't quite see what it is.
>>>>>
>>>>>MTD(f) requires a transposition hashtable and a fail-soft search to function
>>>>>properly. Did you add/change that too?
>>>>
>>>>With hash table, I get this behavior:
>>>>post
>>>>sd 99
>>>>st 99
>>>>go
>>>>Depth Eval Time Nodes PV
>>>>1 48 0 172
>>>>2 0 20 2224 d2d3
>>>>3 35 40 4858
>>>>4 5 321 41505 e2e3
>>>>5 35 711 93485
>>>>6 13 5438 780563 d2d4
>>>>7 30 11878 1638246
>>>>8 18 81509 11563089 e2e4
>>>>move e2e4
>>>>
>>>
>>>I guess you're referring to the empty PV lines?
>>>That's normal, as TSCP doesn't store the PV on a fail high/low. As TSCP is
>>>searching with an open window in the root, you will always have a PV. But with
>>>MTD(f) you're searching with null-windows, so you won't have a PV.
>>>
>>>Did you change stuff in search to deal with pv with fail high/lows?
>>>
>>>In your code below, you're setting a move in the pv array, but that move is
>>>basically random:
>>>
>>> if (f < beta)
>>> greatest = f;
>>> else {
>>> int j;
>>> smallest = f; /* performs a cut-off */
>>> if (smallest >= greatest) {
>>> pv[ply][ply] = gen_dat[depth].m;
>>> for (j = ply + 1; j < pv_length[ply + 1]; ++j)
>>> pv[ply][j] = pv[ply + 1][j];
>>> pv_length[ply] = pv_length[ply + 1];
>>> }
>>> }
>>>
>>>gen_dat[depth].m is a generated move in the movelist, but it is not the
>>>bestmove. It isn't even the first move in the move list as depth is increased on
>>>every iteration.
>>
>>I pulled the pm from the hash table, so the pv move is right (when it is there)
>>but the search still blows chunks of green goo:
>>
>>pvs:
>>post
>>sd 99
>>st 99
>>go
>>Depth Eval Time Nodes PV
>>1 48 0 21 d2d4
>>2 0 0 84 d2d4 d7d5
>>3 35 40 797 d2d4 d7d5 b1c3
>>4 5 80 4313 e2e4 d7d5 f1b5 c8d7 b5d3
>>5 35 241 19619 e2e4 e7e5 d2d4 d7d5 g1f3
>>6 13 872 88472 e2e4 e7e5 d2d4 e5d4 d1d4 g8f6
>>7 30 4517 570969 e2e4 d7d5 e4d5 d8d5 d2d4 d5e4 g1e2 e7e5
>>8 18 30655 3279256 e2e4 b8c6 d2d4 e7e6 b1c3 d7d5 g1f3 g8f6
>>
>>mtdf:
>>post
>>sd 99
>>st 99
>>go
>>Depth Eval Time Nodes PV
>>1 48 0 172
>>2 0 30 2224 d2d4
>>3 35 60 4602
>>4 5 451 41509 e2e4
>>5 35 1362 103001
>>6 13 7741 799578 e2e4
>>7 30 15553 1623933
>>move e2e4
>
>Picking out the hash value and moving the lookup to the end gives a bit better
>result:
>
>/* MTD(f) is an alternative search to pvs */
>int mtdf(int f, int depth)
>{
> int beta;
> int smallest = -INFINITY;
> int greatest = INFINITY;
> TranspositionEntry *entry;
>
> do {
> if (f == smallest)
> beta = f + 1;
> else
> beta = f; /* beta is starting with a first best guess */
> f = AlphaBeta(beta - 1, beta, depth);
>
> if (f < beta)
> greatest = f;
> else {
> int j;
> smallest = f; /* performs a cut-off */
> if (smallest >= greatest) {
> }
> }
> } while (smallest < greatest);
> entry = &transpositionTable[hash & (TRANSPOSITION_SIZE - 1)];
> if (entry->hash == hash) {
> if (entry->move.rawValue != 0 && entry->depth >= depth) {
> pv[ply][0] = entry->move;
> pvLength[ply] = 1;
> }
> }
> return f;
>}
>
>
>post
>sd 99
>st 99
>go
>Depth Eval Time Nodes PV
>1 48 0 172 d2d3
>2 0 20 2224 d2d4
>3 35 40 4602 d2d4
>4 5 320 41509 e2e4
>5 35 751 102983 d2d4
>6 13 5488 799216 e2e4
>7 30 12068 1709733 d2d4
>8 18 82330 11635407 e2e4
>move e2e4
>
>So my next question is, how do you normally populate a hash table with PV nodes,
>since we only get edge values during the search? Do I need to follow the pv
>from hash to hash with a makemove for each succeeding pv node?
>{ICK}
When you fail high in your alphabeta() function, put the move which caused the
fail-high in the hash table.
Andrew
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.