Sherwood, Russell Sunday, July 30, 2017
The Universal Solder?
If you spend any time analyzing with engines and have any ability yourself It becomes obvious that whilst they play close to perfect chess in some positions, they are also clueless in other positions. A traditional example being the King’s Indian Defense, where most engines struggle. An added dimension to this is that different engines excel and misbehave in different types of openings (and/or positions).
Knowing which ones work well (or otherwise) where is a significant potential advantage to the player. Trawling some of the Forums and Webpages can lead to a partial picture, although the problem is that any information gained this way is suspect and could become obsolete in the next version of an engine! If you are wondering why this is so? In simple terms most (stronger) engines are very finely tuned. This tuning means that the engine will play the “best” move in the highest number of positions. An improved version will means the engines plays slightly more positions correctly. The flaw is that, whilst Version 2 might play 94% of positions correctly compared to Version 1 which played 92% of positions in the right manner it does mean core positions – it could well be that 6% of positions more are correct but 4% more are now incorrect. This is general is not an issue, unless, of course, the positions (or more typically themes in positions) are the ones which are important in your opening repertoire. Of course an interesting conclusion from this is that it would be possible to tune an engine for a specific kind of opening but it would probably be weaker in general!
So how can do this ourselves with some degree of certainty? An approach which can deliver dividends is as follows:
- Examine your own opening repertoire
- Identify Key positions in these openings. This can be done from personal knowledge or Opening Books (both Electronic and physical)
- Create an EPD file with these positions and the “Best” moves
- Run a “Beauty contest” with a number of engines to determine which best “understands” these type of positions.
- Use the winner(s) s part of your evaluation strategy in this opening.
Now some may wonder why I believe using the EPD approach is the way forward, rather than Engines Tournaments (either between engines or self-play). The simple reason(s) are : (1) EPD testing is faster and (2) Engine v Engine tournaments can suffer from the results being skewed by either the difference between the two engines or the oddities of an individual engine.
What is worth saying that Engine v Engine can be useful if you can determine which Engine your opponent uses as their main support and openings/lines could then be selected to utilize this information!
Updated Sunday, July 30, 2017 by Russell Sherwood