# Computer Chess Club Archives

## Messages

### Subject: PV verification heuristic

Author: Peter Jacobi

Date: 16:39:10 07/05/98

```Disclaimer: As I wasn't involved with chess programming
for some years, so the topic presented below may have
been resolved for some time.

Having done some games with Comet A90 and Crafty 15.15, I am
somewhat disturbed that I am able to win some games. Even
running the programs on a Pentium 133 shouldn't allow my
mediocre chess (German Rating Number about 1600) to win?!
(15min + 15sec/move time controls)

The same problem will apply to better humans against faster
machines, I assume.

The positions I am able to win, begin with this pattern:
The programs assumes to have some advantage (0.5 ... 2.5),
then following exactly the principal variation displayed, the
program's evaluations goes down, for example because I am able
to regain a "sacrified" piece.

The point I want to raise:
Isn't it possible for the computer to re-check the PV after it
is established, using some (low) fraction of time of the main
analysis.

To elaborate, let the PV be 1.W1 B1, 2.W2 B2, 3.W3 B3,
4.W4, B4 5... - where W1 stands for the first white (computer)
move of the PV, B1 stands for the first black (opponent)
move of the PV, and so on.

Having searched for a level of N ply to get this PV, now the
computer should check the position after 1.W1 B1, 2.W2 for
a level of N-2 plys (or perhaps N-1). This will give one further
ply lookahead. If this doesn't refute (giving a significant
worse eval) the PV, the computer should check the position after
1.W1 B1, 2.W2 B2, 3.W3 for a level of N-3 plys, giving two further
plys lookahead. This can be continued for all or part of the moves
of the PV.

Using this approach, the computer can be somewhat safer against
refutations of his PV, except for the case, that there is a better
black move than 1...B1. For statistical reasons (is this argument
really correct?), more often than not the refutation is not in the
first black move but further down.

And - I don't know yet, what the computer shall do, if it discovers
a refutation.

Peter

```