Sherwood, Russell Thursday, January 18, 2018
Recently I have really got back into playing the King’s Indian Defence. As is well known in CC circles Engines do not understand the KID defence at all, with typical situations having the engine showing a significant plus for white, even when the actual situation is much better for black. There are many reasons why this is the case – three of the main ones being the way engines are tested/tuned, Engine Search Style and the nature of the position itself.
Engine Testing/Tune makes this situation worst as various parameters are set to get the best result in the most possible positions – in other words if I play 10,000 games I want to win as many as possible, changing a parameter might increase the total numbers of wins by a few percentage points but it may well make the evaluation of the engine of “non-standard” positions even worse!
The search style of the engine can make the problem worse. In the simplest sense engines either focus on depth of search or width of search. Engines that focus on depth of search really struggle in the KID as the nature of this search is that a lot of lines are pruned and because the nature of the KID is that many of the lines have sacrifices and deep positional motifs, heavily pruned searching tends to miss this sort of thing, leading to an engine which shows +1.3, +1.4, +1.3, +1.5, 0.25, -0.5, 1.5, falling off a cliff edge as the engine finally “sees” the issue.
The nature of the position itself is one which engines tend to struggle with: Deep and wide attacking lines, sacrifices without immediate benefit…….
So how can we look to better in the KID?
#1 Learn the ideas in the Opening. More so than almost any other opening, it is vital that the player learns the ideas and common manoeuvres seen in the KID. Knowing this will improve the player's sixth sense of which lines to select and which to dismiss.
#2 Select your Engine partner with care. Houdini does better than most of the other major engines In the KID but does not have a learning function (#5)
#3 Build as large a database as possible of games played in the KID.
#4 Build as large an Opening book as possible
#5 Train your engine. A few engines have a “Learning” Mode. What this differs from engine to engine but what we are looking at here is ones that record the best move analysed at a certain depth. This can be useful in itself but comes into its own if paired with a backsliding method. Now we could do this by hand but a more effective way is to utilise game analysis modes seen in most major GUI’s which employ a form a backsliding. This method would allow the engine to build analysis “over the Horizon” based on past games.
#6 Consider using Aquarium’s IDEA function, linked with Infinite Analysis. (which can also be combined with #5!) This allows a very wide search to be utilised to build up a more accurate view of the position(s) over time
#7 Consider Monte-Carlo Analysis – this is an almost random method of building up a view of a position but can help in overcoming the engine’s natural problems.
I know I have enjoyed analysing the KID a lot over recent months but by application of the methods above now have an engine /set of analysis that is approaching a “correct” view of the KID for black.
Updated Thursday, January 18, 2018 by Russell Sherwood