AB & NN - a perfect match?

Sherwood, Russell  Wednesday, May 6, 2020

I’ve been working on a number of articles/books/ideas of late. 

One which is close to fruition is guide to engine analysis, which will be published in written and video format for premium members. Any member wishing to get involved for the proofreading of this, please get in touch.

One question I repeatedly get asked is about the difference between the two engines types and their suitability for positions.

As a way of introduction, when we are selecting a move we are broadly examining three characteristics: Tactical, Positional and Strategic.

Tactical:  referring to a sequence of moves that limits the opponent's options and may result in tangible gain

Positional: referring to a sequence of moves, where the structure of the position is examined and improved (e.g. pawn structure and control of open files)

Strategic: referring to the development of a plan, which goes beyond individual moves or sequence of moves.

There is no absolute difference between the characteristics, they should be considered a continuum. In addition, finding a definitive definition is somewhat difficult!

If we attempt to overlay the main engines around at the moment, we find that most of the main Alpha-Beta engines (of which Stockfish is paramount) are excellent in the Tactical sense, weak1 in the positional sense and generally poor in the Strategic sense1

In addition, it is noticeable that some seemingly positional moves are the result of a deep tactical based search!

Many SF clones tend to be slightly stronger in the positional sense, although often as the expense of some tactical benefits.

Komodo (in AB mode) tends to be slightly stronger positionally and slightly weaker tactically and , on occasion , give the impression of playing strategic moves (in the same way that an exhaustive tactical search throws up positional moves, a deeper search with Komodo can throw up moves that appear strategic in nature)

LC0 tends to play good positional moves and many that are almost strategic nature, although much depends on the network in use and the training regime used (which is another article). It is slightly weaker in terms of tactical play (on typical home hardware) Komodo MCTS shows many of LC0’s characteristics but as its evaluation is still “human”, not quite as well.

An aspiring CC player can utilise this to their advantage. Having recently upgraded to fairly high-end hardware I have been observing SF and Neutral Net engines working in tandem. I have tended to observe (assuming I put 3 MPV on the AB engine) four situations:

  1. Both Engines choose the same moves (although not necessarily the same lines)
  2. The AB Engine (Stockfish) finds a good (tactical) move very quickly and the NN engine significantly longer (typically the move starts towards the bottom of the evaluation and slowly improves its evaluation)
  3. The NN finds a good move , which SF either (a) tends to find after only a very deep search or only after it is “shown” the line.
  4. Both suggest a very different set of moves.

 

If we examine these situations, A happens about 75% of the time, B and C about 10% each and D around 5%. Utilising B and C and checking between the two engines can generate a number of winning “moves”.  D is perhaps the most interesting as this is the area where strategic planning comes into play and is the margin that the highest rated players reside in!

(Note the practical application of this will be covered in the Engine Analysis guide mentioned earlier.

What type of positions do these four scenarios tend to crop up in? We will cover that next time!

 

(1)It is worth noting that “weak” in this sense tends to be mean about 2400 elo!

AnalysisEngines

Updated Wednesday, May 6, 2020 by Russell Sherwood

Welsh Correspondence Chess FederationBritish Correspondence Chess AssociationSchemingMind Internet Correspondence Chess ClubSocial Correspondence Chess AssociationNational Correspondence Chess ClubWelsh Chess UnionInternational Correspondence Chess Association