Hi Matthew,

"Chess thinking", as you say, really occurs in two forms in a chess program.

1. Search
2. Evaluation

At some point in time, as the program generates POSITIONS from its move generator, it must stop, and, WITHOUT searching, evaluate the position.

The is called a LEAF NODE EVALUATION, and, at best, it is a crap shoot. The dominant form of the evaluation is material, and because no search is performed at this stage, something deadly can be one or two moves away, and the program does not know it.

The good news: all leaf nodes are in the distant (8 plies, 10 plies, or more) in the future, so, statistically, ANY ONE SINGLE NODE will most likely NOT be a factor in the outcome of the game.

More good news: millions of leaf nodes are evaluated, most are junk and discarded, and this filtering process means that only the "balanced" positions survive to be passed down further into the tree.

So, "chess thinking" is really an idiotic form of trying millions of things that don't work, distilling this down into just one PRINCIPLE VARIATION, which is the analysis you see as the search builds.

The "PV" is the result of all of the lead node evals being passed back and forth through the ALPHA BETA search. The APLHA side always wants to play the move leading to the biggest score for it, and the BETA side always wants to play the move leading to the smallest score for the ALPHA side.

In this way, one side makes the "strongest move", and the other side makes the best reply to it, and so on.

Where the "intelligence" comes in is in the leaf mode evaluation routine.

There are way to encode positions that are known wins/losses/draws so that the leaf node eval will OVERRIDE the material score.

This takes intelligence.

For example, if you have 1 knight, and your opponent has just his king, you would not want that to be scored as +3 pawns for the knight (say + 300) since it is a dead draw.

I am surprised at how many commercial programs will start to search in such a K+N vs. K position and return a +300 score and actually try to win.

Vortex makes no such errors. In fact, Vortex knows NN vs. K is not a +600 score since so many NN vs. K positions are drawn. It will not "dismiss" the position as a mere draw, since a falible player can mess up the ending and walk into a mate.

But, Vortex would prefer two unconnected passed pawns on the a- and j-files rather than having 2 knights, since its evaluation function has intelligence identifying which types of endgames lead to wins.

Vortex can identify ANY position with X pawns vs. Y pawns as a win, even with a 1-ply search! This took a great deal of intelligent coding!! It has a "pawn evaluator" that is pretty much always correct. So, Vortex will sometimes swap pieces like crazy as the endgame approaches, only to be able to take you into an incredibly complex king and pawn ending where it will win with no trouble.

There are many such "patterns" that make up its intelligence. It knows R + P endings well, it knows Bishop + wrong Rook's Pawn draws WITHOUT having to search (meaning the leaf node eval will handle it properly in an instant) and many such thematic ideas that will overpower the material evaluator.

Once I hook up the 5-piece Gothic Chess endgame databases to it in a RAM buffer, its play will be amazing as the endgame approaches.

I am not there yet though.